Merge pull request #120 from TonyJiangWJ/master

添加 PaddleOCR 功能 by TonyJiangWJ

app/build.gradle.kts

@@ -109,6 +109,9 @@ dependencies /* Unclassified */ {
     // OpenCV
     implementation(project(":libs:org.opencv-4.5.5"))
 
+    // PaddleOCR
+    implementation(project(":libs:paddleocr"))
+
     // Android Job
     implementation(project(":libs:android-job-simplified-1.4.3"))
 

app/src/main/assets/init.js

@@ -485,7 +485,7 @@ Object.assign(this, {
 
                 /* ! ocr < images */
                 /* ! ocr < files */
-                [ 'ocr' ],
+                [ 'ocr', 'paddle_ocr' ],
 
                 /* Safe to put last regardless of the order, no guarantee ;). */
                 [ 'floaty', 'storages', 'device', 'recorder', 'toast' ],

app/src/main/assets/modules/__paddle_ocr__.js

@@ -0,0 +1,120 @@
+// noinspection UnnecessaryLocalVariableJS,JSUnusedLocalSymbols
+
+/* Overwritten protection. */
+
+let { images } = global;
+
+/**
+ * @param {org.autojs.autojs.runtime.ScriptRuntime} scriptRuntime
+ * @param {org.mozilla.javascript.Scriptable | global} scope
+ * @return {Internal.Ocr}
+ */
+module.exports = function (scriptRuntime, scope) {
+
+    const rtOcr = scriptRuntime.paddleOCR;
+
+    let _ = {
+        OcrCtor: (/* @IIFE */ () => {
+            /**
+             * @implements Internal.Ocr
+             */
+            const OcrCtor = function () {
+
+                /** @global */
+                const ocr = function (img, options) {
+                    if (typeof arguments[0] === 'string') {
+                        let img = images.read(/* path = */ arguments[0]);
+                        if (img === null) {
+                            throw TypeError(`Invalid image of path "${arguments[0]}" for ocr(img, options?)`);
+                        }
+                        return OcrCtor.prototype.recognizeText(img.oneShot(), options);
+                    }
+                    return OcrCtor.prototype.recognizeText.apply(OcrCtor.prototype, arguments);
+                };
+
+                return Object.assign(ocr, OcrCtor.prototype);
+            };
+
+            OcrCtor.prototype = {
+                constructor: OcrCtor,
+                recognizeText(img, options) {
+                    if (typeof arguments[0] === 'string') {
+                        let img = images.read(/* path = */ arguments[0]);
+                        if (img === null) {
+                            throw TypeError(`Invalid image of path "${arguments[0]}" for ocr.recognizeText(img, options?)`);
+                        }
+                        return this.recognizeText(img.oneShot(), options);
+                    }
+                    if (_.shouldTakenAsRegion(arguments[1])) {
+                        return this.recognizeText(img, { region: arguments[1] });
+                    }
+                    let opt = options || {};
+                    let region = opt.region;
+                    if (region === null) {
+                        return [];
+                    }
+                    let cpuThreadNum = opt.cpuThreadNum || 4
+                    let useSlim = opt.useSlim
+                    if (useSlim === undefined) {
+                       // 默认使用轻量化模型
+                       useSlim = true
+                    }
+                    if (region === undefined) {
+                        return Array.from(rtOcr.recognizeText(img, cpuThreadNum, useSlim));
+                    }
+                    let results = Array.from(rtOcr.recognizeText(images.clip(img, region).oneShot()));
+                    img.shoot();
+                    return results;
+                },
+                detect(img, options) {
+                    if (typeof arguments[0] === 'string') {
+                        let img = images.read(/* path = */ arguments[0]);
+                        if (img === null) {
+                            throw TypeError(`Invalid image of path "${arguments[0]}" for ocr.detect(img, options?)`);
+                        }
+                        return this.detect(img.oneShot(), options);
+                    }
+                    if (_.shouldTakenAsRegion(arguments[1])) {
+                        return this.detect(img, { region: arguments[1] });
+                    }
+                    let opt = options || {};
+                    let region = opt.region;
+                    if (region === null) {
+                        return [];
+                    }
+                    let cpuThreadNum = opt.cpuThreadNum || 4
+                    let useSlim = opt.useSlim
+                    if (useSlim === undefined) {
+                       // 默认使用轻量化模型
+                       useSlim = true
+                    }
+                    /**
+                     * @type {org.autojs.autojs.runtime.api.OcrResult[]}
+                     */
+                    let resultList = rtOcr.detect(region !== undefined ? images.clip(img, region).oneShot() : img, cpuThreadNum, useSlim).toArray();
+                    if (!isNullish(region)) {
+                        resultList.forEach((result) => {
+                            let rect = images.buildRegion(img, region);
+                            result.bounds.offset(rect.x, rect.y);
+                        });
+                    }
+                    return Array.from(resultList);
+                },
+            };
+
+            return OcrCtor;
+        })(),
+        shouldTakenAsRegion(o) {
+            return o instanceof org.opencv.core.Rect
+                || o instanceof android.graphics.Rect
+                || Array.isArray(o);
+        },
+    };
+
+    /**
+     * @type {Internal.Ocr}
+     */
+    const ocr = new _.OcrCtor();
+
+    return ocr;
+};

app/src/main/assets/sample/OCR/PaddleOCR内置API.js

@@ -0,0 +1,15 @@
+/**
+ * @author TonyJiangWJ
+ */
+console.show()
+// 指定是否用精简版模型 速度较快 默认为true
+let useSlim = false
+// cpu线程数量，实际好像没啥作用
+let cpuThreadNum = 4
+let start = new Date()
+let img = images.read('test.png')
+let results = paddle_ocr.detect(img, cpuThreadNum, useSlim)
+toastLog('识别结束, 耗时：' + (new Date() - start) + 'ms')
+log('识别结果：' + JSON.stringify(Array.from(results).map(result => ({ label: result.label, confidence: result.confidence, bounds: result.bounds }))))
+// 回收图片
+img.recycle()

app/src/main/assets/sample/OCR/PaddleOCR原始类.js

@@ -0,0 +1,39 @@
+/**
+ * @author TonyJiangWJ
+ */
+importClass(com.baidu.paddle.lite.ocr.Predictor)
+
+console.show()
+// 指定是否用精简版模型 速度较快
+let useSlim = false
+// 创建检测器
+let predictor = new Predictor()
+// predictor.cpuThreadNum = 4 //可以自定义使用CPU的线程数
+// predictor.checkModelLoaded = false // 可以自定义是否需要校验模型是否成功加载 默认开启 使用内置Base64图片进行校验 识别测试文本来校验模型是否加载成功
+// 初始化模型 首次运行时会比较耗时
+let loading = threads.disposable()
+// 建议在新线程中初始化模型
+threads.start(function () {
+  loading.setAndNotify(predictor.init(context, useSlim))
+  // loading.setAndNotify(predictor.init(context)) 为默认不使用精简版
+  // 内置默认 modelPath 为 models/ocr_v3_for_cpu，初始化自定义模型请写绝对路径否则无法获取到
+  // 内置默认 labelPath 为 labels/ppocr_keys_v1.txt
+  // let modelPath = files.path('./models/customize') // 指定自定义模型路径
+  // let labelPath = files.path('./models/customize') // 指定自定义label路径
+  // 使用自定义模型时det rec cls三个模型文件名称需要手动指定
+  // predictor.detModelFilename = 'det_opt.nb'
+  // predictor.recModelFilename = 'rec_opt.nb'
+  // predictor.clsModelFilename = 'cls_opt.nb'
+  // loading.setAndNotify(predictor.init(context, modelPath, labelPath))
+})
+let loadSuccess = loading.blockedGet()
+toastLog('加载模型结果：' + loadSuccess)
+let start = new Date()
+let img = images.read('test.png')
+let results = predictor.runOcr(img.getBitmap())
+toastLog('识别结束, 耗时：' + (new Date() - start) + 'ms')
+log('识别结果：' + JSON.stringify(Array.from(results).map(result => ({ label: result.label, confidence: result.confidence, bounds: result.bounds }))))
+// 释放模型 用于释放native内存 非必需
+// predictor.releaseModel()
+// 回收图片
+img.recycle()

app/src/main/assets/sample/OCR/PaddleOCR截图识别.js

@@ -0,0 +1,197 @@
+/**
+ * create by TonyJiangWJ
+ */
+let currentEngine = engines.myEngine()
+let runningEngines = engines.all()
+let currentSource = currentEngine.getSource() + ''
+if (runningEngines.length > 1) {
+  runningEngines.forEach(compareEngine => {
+    let compareSource = compareEngine.getSource() + ''
+    if (currentEngine.id !== compareEngine.id && compareSource === currentSource) {
+      // 强制关闭同名的脚本
+      compareEngine.forceStop()
+    }
+  })
+}
+
+
+if (!requestScreenCapture()) {
+  toastLog('请求截图权限失败')
+  exit()
+}
+
+sleep(1000)
+
+importClass(com.baidu.paddle.lite.ocr.Predictor)
+
+// 指定是否用精简版模型 速度较快
+let useSlim = true
+// 创建检测器
+let predictor = new Predictor()
+// predictor.cpuThreadNum = 4 // 可以自定义使用CPU的线程数
+// predictor.checkModelLoaded = false // 可以自定义是否需要校验模型是否成功加载 默认开启 使用内置Base64图片进行校验 识别测试文本来校验模型是否加载成功
+// 初始化模型 首次运行时会比较耗时
+let loading = threads.disposable()
+// 建议在新线程中初始化模型
+threads.start(function () {
+  loading.setAndNotify(predictor.init(context, useSlim))
+  // loading.setAndNotify(predictor.init(context)) 为默认不使用精简版
+  // 内置默认 modelPath 为 models/ocr_v3_for_cpu，初始化自定义模型请写绝对路径否则无法获取到
+  // 内置默认 labelPath 为 labels/ppocr_keys_v1.txt
+  // let modelPath = files.path('./models/customize') // 指定自定义模型路径
+  // let labelPath = files.path('./models/customize') // 指定自定义label路径
+  // 使用自定义模型时det rec cls三个模型文件名称需要手动指定
+  // predictor.detModelFilename = 'det_opt.nb'
+  // predictor.recModelFilename = 'rec_opt.nb'
+  // predictor.clsModelFilename = 'cls_opt.nb'
+  // loading.setAndNotify(predictor.init(context, modelPath, labelPath))
+})
+let loadSuccess = loading.blockedGet()
+if (!loadSuccess) {
+  toastLog('初始化ocr失败')
+  exit()
+}
+// 识别结果和截图信息
+let result = []
+let img = null
+let running = true
+let capturing = true
+
+/**
+ * 截图并识别OCR文本信息
+ */
+function captureAndOcr() {
+  capturing = true
+  img && img.recycle()
+  img = captureScreen()
+  if (!img) {
+    toastLog('截图失败')
+  }
+  let start = new Date()
+  result = predictor.runOcr(img.getBitmap())
+  toastLog('耗时' + (new Date() - start) + 'ms')
+  capturing = false
+}
+
+captureAndOcr()
+
+// 获取状态栏高度
+let offset = -getStatusBarHeightCompat()
+
+// 绘制识别结果
+let window = floaty.rawWindow(
+  <canvas id="canvas" layout_weight="1" />
+);
+
+// 设置悬浮窗位置
+ui.post(() => {
+  window.setPosition(0, offset)
+  window.setSize(device.width, device.height)
+  window.setTouchable(false)
+})
+
+// 操作按钮
+let clickButtonWindow = floaty.rawWindow(
+  <vertical>
+    <button id="captureAndOcr" text="截图识别" />
+    <button id="closeBtn" text="退出" />
+  </vertical>
+);
+ui.run(function () {
+  clickButtonWindow.setPosition(device.width / 2 - ~~(clickButtonWindow.getWidth() / 2), device.height * 0.65)
+})
+
+// 点击识别
+clickButtonWindow.captureAndOcr.click(function () {
+  result = []
+  ui.run(function () {
+    clickButtonWindow.setPosition(device.width, device.height)
+  })
+  setTimeout(() => {
+    captureAndOcr()
+    ui.run(function () {
+      clickButtonWindow.setPosition(device.width / 2 - ~~(clickButtonWindow.getWidth() / 2), device.height * 0.65)
+    })
+  }, 500)
+})
+
+// 点击关闭
+clickButtonWindow.closeBtn.click(function () {
+  exit()
+})
+
+let Typeface = android.graphics.Typeface
+let paint = new Paint()
+paint.setStrokeWidth(1)
+paint.setTypeface(Typeface.DEFAULT_BOLD)
+paint.setTextAlign(Paint.Align.LEFT)
+paint.setAntiAlias(true)
+paint.setStrokeJoin(Paint.Join.ROUND)
+paint.setDither(true)
+window.canvas.on('draw', function (canvas) {
+  if (!running || capturing) {
+    return
+  }
+  // 清空内容
+  canvas.drawColor(0xFFFFFF, android.graphics.PorterDuff.Mode.CLEAR)
+  if (result && result.length > 0) {
+    for (let i = 0; i < result.length; i++) {
+      let ocrResult = result[i]
+      drawRectAndText(ocrResult.label, ocrResult.bounds, '#00ff00', canvas, paint)
+    }
+  }
+})
+
+setInterval(() => { }, 10000)
+events.on('exit', () => {
+  // 标记停止 避免canvas导致闪退
+  running = false
+  // 撤销监听
+  window.canvas.removeAllListeners()
+  // 回收图片
+  img && img.recycle()
+})
+
+/**
+ * 绘制文本和方框
+ *
+ * @param {*} desc
+ * @param {*} rect
+ * @param {*} colorStr
+ * @param {*} canvas
+ * @param {*} paint
+ */
+function drawRectAndText (desc, rect, colorStr, canvas, paint) {
+  let color = colors.parseColor(colorStr)
+
+  paint.setStrokeWidth(1)
+  paint.setStyle(Paint.Style.STROKE)
+  // 反色
+  paint.setARGB(255, 255 - (color >> 16 & 0xff), 255 - (color >> 8 & 0xff), 255 - (color & 0xff))
+  canvas.drawRect(rect, paint)
+  paint.setARGB(255, color >> 16 & 0xff, color >> 8 & 0xff, color & 0xff)
+  paint.setStrokeWidth(1)
+  paint.setTextSize(20)
+  paint.setStyle(Paint.Style.FILL)
+  canvas.drawText(desc, rect.left, rect.top, paint)
+  paint.setTextSize(10)
+  paint.setStrokeWidth(1)
+  paint.setARGB(255, 0, 0, 0)
+}
+
+/**
+ * 获取状态栏高度
+ *
+ * @returns
+ */
+function getStatusBarHeightCompat () {
+  let result = 0
+  let resId = context.getResources().getIdentifier("status_bar_height", "dimen", "android")
+  if (resId > 0) {
+    result = context.getResources().getDimensionPixelOffset(resId)
+  }
+  if (result <= 0) {
+    result = context.getResources().getDimensionPixelOffset(R.dimen.dimen_25dp)
+  }
+  return result
+}

app/src/main/java/org/autojs/autojs/runtime/ScriptRuntime.java

@@ -40,6 +40,7 @@ import org.autojs.autojs.runtime.api.Floaty;
 import org.autojs.autojs.runtime.api.Images;
 import org.autojs.autojs.runtime.api.Media;
 import org.autojs.autojs.runtime.api.MlKitOCR;
+import org.autojs.autojs.runtime.api.PaddleOCR;
 import org.autojs.autojs.runtime.api.Plugins;
 import org.autojs.autojs.runtime.api.ProcessShell;
 import org.autojs.autojs.runtime.api.ScreenMetrics;
@@ -243,6 +244,9 @@ public class ScriptRuntime {
     @ScriptVariable
     public final MlKitOCR mlKitOCR;
 
+    @ScriptVariable
+    public final PaddleOCR paddleOCR;
+
     private static WeakReference<Context> applicationContext;
     private final Map<String, Object> mProperties = new ConcurrentHashMap<>();
     private AbstractShell mRootShell;
@@ -278,6 +282,7 @@ public class ScriptRuntime {
         plugins = new Plugins(context, this);
 
         mlKitOCR = new MlKitOCR();
+        paddleOCR = new PaddleOCR();
     }
 
     public void init() {
@@ -554,6 +559,7 @@ public class ScriptRuntime {
         ignoresException(this::recycleShell);
         ignoresException(images::releaseScreenCapturer);
         ignoresException(mlKitOCR::release);
+        ignoresException(paddleOCR::release);
         ignoresException(sensors::unregisterAll);
         ignoresException(timers::recycle);
         ignoresException(ui::recycle);

app/src/main/java/org/autojs/autojs/runtime/api/PaddleOCR.java

@@ -0,0 +1,85 @@
+package org.autojs.autojs.runtime.api;
+
+import android.graphics.Bitmap;
+import android.os.Looper;
+import android.util.Log;
+
+import com.baidu.paddle.lite.ocr.OcrResult;
+import com.baidu.paddle.lite.ocr.Predictor;
+import org.autojs.autojs.app.GlobalAppContext;
+import org.autojs.autojs.concurrent.VolatileDispose;
+import org.autojs.autojs.core.image.ImageWrapper;
+
+import java.util.Collections;
+import java.util.List;
+
+/**
+ * @author TonyJiangWJ
+ * @since 2023-08-06
+ */
+public class PaddleOCR {
+    private final Predictor mPredictor = new Predictor();
+
+    public synchronized boolean init(boolean useSlim) {
+        if (!mPredictor.isLoaded || useSlim != mPredictor.isUseSlim()) {
+            if (Looper.getMainLooper() == Looper.myLooper()) {
+                VolatileDispose<Boolean> result = new VolatileDispose<>();
+                new Thread(() -> {
+                    result.setAndNotify(mPredictor.init(GlobalAppContext.get(), useSlim));
+                }).start();
+                return result.blockedGet();
+            } else {
+                return mPredictor.init(GlobalAppContext.get(), useSlim);
+            }
+        }
+        return mPredictor.isLoaded;
+    }
+
+    public void release() {
+        mPredictor.releaseModel();
+    }
+
+    public List<OcrResult> detect(ImageWrapper image, int cpuThreadNum, boolean useSlim) {
+        if (image == null) {
+            return Collections.emptyList();
+        }
+        Bitmap bitmap = image.getBitmap();
+        if (bitmap.isRecycled()) {
+            return Collections.emptyList();
+        }
+        if (mPredictor.cpuThreadNum != cpuThreadNum) {
+            mPredictor.releaseModel();
+            mPredictor.cpuThreadNum = cpuThreadNum;
+        }
+        init(useSlim);
+        return mPredictor.runOcr(bitmap);
+    }
+
+    public List<OcrResult> detect(ImageWrapper image, int cpuThreadNum) {
+        return detect(image, cpuThreadNum, true);
+    }
+
+    public List<OcrResult> detect(ImageWrapper image) {
+        return detect(image, 4, true);
+    }
+
+    public String[] recognizeText(ImageWrapper image, int cpuThreadNum, boolean useSlim) {
+        List<OcrResult> words_result = detect(image, cpuThreadNum, useSlim);
+        Collections.sort(words_result);
+        String[] outputResult = new String[words_result.size()];
+        for (int i = 0; i < words_result.size(); i++) {
+            outputResult[i] = words_result.get(i).getLabel();
+            // show LOG in Logcat panel
+            Log.i("outputResult", outputResult[i]);
+        }
+        return outputResult;
+    }
+
+    public String[] recognizeText(ImageWrapper image, int cpuThreadNum) {
+        return recognizeText(image, cpuThreadNum, true);
+    }
+
+    public String[] recognizeText(ImageWrapper image) {
+        return recognizeText(image, 4, true);
+    }
+}

app/src/main/res/raw/licenses.xml

@@ -101,4 +101,10 @@
         <url>http://www.lingala.net/zip4j/</url>
         <license>Apache Software License 2.0</license>
     </notice>
+    <notice>
+        <name>paddle</name>
+        <copyright>Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved</copyright>
+        <url>https://github.com/paddlepaddle/paddle</url>
+        <license>Apache Software License 2.0</license>
+    </notice>
 </notices>

libs/paddleocr/.gitignore

@@ -0,0 +1,4 @@
+/build
+.cxx/
+cache/
+OpenCV/

libs/paddleocr/PaddleLite/cxx/include/paddle_api.h

@@ -0,0 +1,611 @@
+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+/*
+ * This file defines PaddlePredictor, the api for lite. It supports multiple
+ * hardware including ARM, X86, OpenCL, CUDA and so on.
+ */
+
+#ifndef PADDLE_LITE_API_H_  // NOLINT
+#define PADDLE_LITE_API_H_
+#include <functional>
+#include <map>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+#include "paddle_place.h"  // NOLINT
+
+namespace paddle {
+namespace lite_api {
+
+using shape_t = std::vector<int64_t>;
+using lod_t = std::vector<std::vector<uint64_t>>;
+
+enum class LiteModelType { kProtobuf = 0, kNaiveBuffer, UNK };
+// Methods for allocating L3Cache on Arm platform
+enum class L3CacheSetMethod {
+  kDeviceL3Cache = 0,  // Use the system L3 Cache size, best performance.
+  kDeviceL2Cache = 1,  // Use the system L2 Cache size, trade off performance
+                       // with less memory consumption.
+  kAbsolute = 2,       // Use the external setting.
+  // kAutoGrow = 3,   // Not supported yet, least memory consumption.
+};
+
+// return true if current device supports OpenCL model
+LITE_API bool IsOpenCLBackendValid(bool check_fp16_valid = false);
+
+// return current opencl device type,
+// if opencl not enabled or IsOpenCLBackendValid return false, it will return -1
+// UNKNOWN:0, QUALCOMM_ADRENO:1, ARM_MALI:2, IMAGINATION_POWERVR:3, OTHERS:4,
+LITE_API int GetOpenCLDeviceType();
+
+struct LITE_API Tensor {
+  explicit Tensor(void* raw);
+  explicit Tensor(const void* raw);
+
+  void Resize(const shape_t& shape);
+
+  /// Readonly data.
+  template <typename T>
+  const T* data() const;
+
+  template <typename T>
+  T* mutable_data(TargetType type = TargetType::kHost) const;
+
+  void* mutable_metal_data(void* ptr) const;
+
+  // Share external memory. Note: ensure that the data pointer is in a valid
+  // state
+  // during the prediction process.
+  void ShareExternalMemory(void* data, size_t memory_size, TargetType target);
+
+  template <typename T, TargetType type = TargetType::kHost>
+  void CopyFromCpu(const T* data);
+
+  template <typename T>
+  void CopyToCpu(T* data) const;
+  /// Shape of the tensor.
+  shape_t shape() const;
+  TargetType target() const;
+  PrecisionType precision() const;
+  void SetPrecision(PrecisionType precision);
+
+  // LoD of the tensor
+  lod_t lod() const;
+
+  // Set LoD of the tensor
+  void SetLoD(const lod_t& lod);
+  bool IsInitialized() const;
+
+ private:
+  void* raw_tensor_;
+};
+
+/// The PaddlePredictor defines the basic interfaces for different kinds of
+/// predictors.
+class LITE_API PaddlePredictor {
+ public:
+  PaddlePredictor() = default;
+
+  /// Get i-th input.
+  virtual std::unique_ptr<Tensor> GetInput(int i) = 0;
+
+  /// Get i-th output.
+  virtual std::unique_ptr<const Tensor> GetOutput(int i) const = 0;
+
+  virtual void Run() = 0;
+  virtual std::shared_ptr<PaddlePredictor> Clone() = 0;
+  virtual std::shared_ptr<PaddlePredictor> Clone(
+      const std::vector<std::string>& var_names) = 0;
+
+  virtual std::string GetVersion() const = 0;
+
+  // Get input names
+  virtual std::vector<std::string> GetInputNames() = 0;
+  // Get output names
+  virtual std::vector<std::string> GetOutputNames() = 0;
+  // Get output names
+  virtual std::vector<std::string> GetParamNames();
+
+  /// Release all tmp tensor to compress the size of the memory pool.
+  virtual bool TryShrinkMemory() = 0;
+
+  // Get Input by name
+  virtual std::unique_ptr<Tensor> GetInputByName(const std::string& name) = 0;
+
+  /// Get a readonly tensor, return null if no one called `name` exists.
+  virtual std::unique_ptr<const Tensor> GetTensor(
+      const std::string& name) const = 0;
+  /// Get a mutable tensor, return null if on one called `name` exists
+  /// internal infereces API, not recommanded.
+  virtual std::unique_ptr<Tensor> GetMutableTensor(const std::string& name);
+
+  /// Persist the optimized model to disk. This API is only supported by
+  /// CxxConfig, and the persisted model can be reused for MobileConfig.
+  virtual void SaveOptimizedModel(
+      const std::string& model_dir,
+      LiteModelType model_type = LiteModelType::kProtobuf,
+      bool record_info = false);
+
+  virtual ~PaddlePredictor() = default;
+
+ protected:
+  int threads_{1};
+  lite_api::PowerMode mode_{lite_api::LITE_POWER_NO_BIND};
+};
+
+/// Base class for all the configs.
+class LITE_API ConfigBase {
+  std::string model_dir_;
+  int threads_{1};
+  PowerMode mode_{LITE_POWER_NO_BIND};
+  // gpu opencl
+  CLTuneMode opencl_tune_mode_{CL_TUNE_NONE};
+  std::string opencl_bin_path_{""};
+  std::string opencl_bin_name_{""};
+  CLPrecisionType opencl_precision_{CL_PRECISION_AUTO};
+  // Where to cache the npu/xpu/rknpu/apu offline model to the binary files
+  std::string subgraph_model_cache_dir_{""};
+  // Set the cached npu/xpu/rknpu/apu offline model from the buffers
+  std::map<std::string, std::pair<std::vector<char>, std::vector<char>>>
+      subgraph_model_cache_buffers_{};
+  // The selected NNAdapter devices to build and run the model.
+  std::vector<std::string> nnadapter_device_names_{};
+  // The NNAdapter context properties for device configuration, model
+  // compilation and execution
+  std::string nnadapter_context_properties_{};
+  int (*nnadapter_context_callback_)(int event_id,
+                                     void* user_data){nullptr};  // NOLINT
+  // The directory to find and store the compiled NNAdapter models.
+  std::string nnadapter_model_cache_dir_{""};
+  // Dynamic shapes of the NNAdapter model
+  std::map<std::string, std::vector<std::vector<int64_t>>>
+      nnadapter_dynamic_shape_info_;
+  // The buffers for loading the compiled NNAdapter models from memory.
+  std::map<std::string, std::vector<char>> nnadapter_model_cache_buffers_{};
+  int device_id_{0};
+  int x86_math_num_threads_ = 1;
+
+  std::string metal_path_;
+  bool metal_use_mps_{false};
+  bool metal_use_aggressive_{false};
+  void* metal_device_{nullptr};
+  bool metal_use_memory_reuse_{false};
+
+  std::vector<std::string> discarded_passes_{};
+
+ public:
+  explicit ConfigBase(PowerMode mode = LITE_POWER_NO_BIND, int threads = 1);
+  // set Model_dir
+  void set_model_dir(const std::string& x) { model_dir_ = x; }
+  const std::string& model_dir() const { return model_dir_; }
+  // set Thread
+  void set_threads(int threads);
+  int threads() const { return threads_; }
+  // set Power_mode
+  void set_power_mode(PowerMode mode);
+  PowerMode power_mode() const { return mode_; }
+
+  /// \brief Set path and file name of generated OpenCL compiled kernel binary.
+  ///
+  /// If you use GPU of specific soc, using OpenCL binary will speed up the
+  /// initialization.
+  ///
+  /// \param path  Path that OpenCL compiled kernel binay file stores in. Make
+  /// sure the path exist and you have Read&Write permission.
+  /// \param name  File name of OpenCL compiled kernel binay.
+  /// \return void
+  void set_opencl_binary_path_name(const std::string& path,
+                                   const std::string& name);
+
+  /// \brief Set path and file name of generated OpenCL algorithm selecting
+  /// file.
+  ///
+  /// If you use GPU of specific soc, using OpenCL binary will speed up the
+  /// running time in most cases. But the first running for algorithm selecting
+  /// is timg-costing.
+  ///
+  /// \param tune_mode  Set a tune mode:
+  ///        CL_TUNE_NONE: turn off
+  ///        CL_TUNE_RAPID: find the optimal algorithm in a rapid way(less
+  ///        time-cost)
+  ///        CL_TUNE_NORMAL: find the optimal algorithm in a noraml
+  ///        way(suggestion)
+  ///        CL_TUNE_EXHAUSTIVE: find the optimal algorithm in a exhaustive
+  ///        way(most time-costing)
+  /// \param path  Path that OpenCL algorithm selecting file stores in. Make
+  /// sure the path exist and you have Read&Write permission.
+  /// \param name  File name of OpenCL algorithm selecting file.
+  /// \param lws_repeats  Repeat number for find the optimal local work size .
+  /// \return void
+  void set_opencl_tune(CLTuneMode tune_mode = CL_TUNE_NONE,
+                       const std::string& path = "",
+                       const std::string& name = "",
+                       size_t lws_repeats = 4);
+
+  /// \brief Set runtime precision on GPU using OpenCL backend.
+  ///
+  /// \param p
+  ///          CL_PRECISION_AUTO: first fp16 if valid, default
+  ///          CL_PRECISION_FP32: force fp32
+  ///          CL_PRECISION_FP16: force fp16
+  /// \return void
+  void set_opencl_precision(CLPrecisionType p = CL_PRECISION_AUTO);
+
+  // set subgraph_model_dir
+  void set_subgraph_model_cache_dir(std::string subgraph_model_cache_dir) {
+    subgraph_model_cache_dir_ = subgraph_model_cache_dir;
+  }
+  const std::string& subgraph_model_cache_dir() const {
+    return subgraph_model_cache_dir_;
+  }
+  void set_subgraph_model_cache_buffers(const std::string& key,
+                                        const std::vector<char>& cfg,
+                                        const std::vector<char>& bin);
+  const std::map<std::string, std::pair<std::vector<char>, std::vector<char>>>&
+  subgraph_model_cache_buffers() const {
+    return subgraph_model_cache_buffers_;
+  }
+  // Check if the NNAdapter device is valid.
+  bool check_nnadapter_device_name(const std::string& device_name);
+  // Choose the NNAdapter devices to build and run the model.
+  void set_nnadapter_device_names(
+      const std::vector<std::string>& device_names) {
+    nnadapter_device_names_ = device_names;
+  }
+  const std::vector<std::string>& nnadapter_device_names() const {
+    return nnadapter_device_names_;
+  }
+  // Set the context properties by key-value map for NNAdapter device
+  // configuration, model compilation and execution
+  // Such as "HUAWEI_ASCEND_NPU_SELECTED_DEVICE_IDS=0;"
+  void set_nnadapter_context_properties(const std::string& context_properties) {
+    nnadapter_context_properties_ = context_properties;
+  }
+  const std::string& nnadapter_context_properties() const {
+    return nnadapter_context_properties_;
+  }
+  // Set nnadapter_context_callback for NNAdapter device to get runtime
+  // parameters.
+  // For example:
+  // cudaStream_t cuda_stream;
+  // cudaStreamCreate(&cuda_stream);
+  // int nnadapter_context_callback(int event_id, void* user_data) {
+  //   if (event_id == 0x0100) {
+  //     *(std::reinterpret_cast<cudaStream_t*>(user_data)) = cuda_stream;
+  //   }
+  //   return 0;
+  // }
+  void set_nnadapter_context_callback(
+      int (*nnadapter_context_callback)(int event_id, void* user_data)) {
+    nnadapter_context_callback_ = nnadapter_context_callback;
+  }
+  int (*nnadapter_context_callback() const)(int event_id,  // NOLINT
+                                            void* user_data) {
+    return nnadapter_context_callback_;
+  }
+
+  // Enable caching and set the directory to search and store the compiled
+  // NNAdapter models in the file system.
+  void set_nnadapter_model_cache_dir(const std::string& model_cache_dir) {
+    nnadapter_model_cache_dir_ = model_cache_dir;
+  }
+  const std::string& nnadapter_model_cache_dir() const {
+    return nnadapter_model_cache_dir_;
+  }
+  // Set dynamic shapes for building models
+  void set_nnadapter_dynamic_shape_info(
+      const std::map<std::string, std::vector<std::vector<int64_t>>>&
+          nnadapter_dynamic_shape_info) {
+    nnadapter_dynamic_shape_info_ = nnadapter_dynamic_shape_info;
+  }
+  const std::map<std::string, std::vector<std::vector<int64_t>>>&
+  nnadapter_dynamic_shape_info() const {
+    return nnadapter_dynamic_shape_info_;
+  }
+  // Set the buffers for loading the compiled NNAdapter models from memory.
+  void set_nnadapter_model_cache_buffers(
+      const std::string& model_cache_token,
+      const std::vector<char>& model_cache_buffer);
+  const std::map<std::string, std::vector<char>>&
+  nnadapter_model_cache_buffers() const {
+    return nnadapter_model_cache_buffers_;
+  }
+  // set Device ID
+  void set_device_id(int device_id) { device_id_ = device_id; }
+  int get_device_id() const { return device_id_; }
+  // set x86_math_num_threads
+  void set_x86_math_num_threads(int threads);
+  int x86_math_num_threads() const;
+
+  void set_metal_lib_path(const std::string& path);
+  void set_metal_use_mps(bool flag);
+  void set_metal_use_aggressive(bool flag);
+  void set_metal_device(void* device);
+  void set_metal_use_memory_reuse(bool flag);
+
+  std::string metal_lib_path() const { return metal_path_; }
+  bool metal_use_mps() const { return metal_use_mps_; }
+  bool metal_use_aggressive() const { return metal_use_aggressive_; }
+  void* metal_device() const { return metal_device_; }
+  bool metal_use_memory_reuse() const { return metal_use_memory_reuse_; }
+
+  void add_discarded_pass(const std::string pass);
+  const std::vector<std::string> get_discarded_passes() const {
+    return discarded_passes_;
+  }
+};
+
+class LITE_API CxxModelBuffer {
+ public:
+  CxxModelBuffer(const char* program_buffer,
+                 size_t program_buffer_size,
+                 const char* params_buffer,
+                 size_t params_buffer_size);
+  CxxModelBuffer(std::string&& program_buffer, std::string&& params_buffer);
+  const std::string& get_program() const;
+  const std::string& get_params() const;
+  bool is_empty() const;
+
+  CxxModelBuffer() = default;
+  CxxModelBuffer(const CxxModelBuffer&) = delete;
+
+ private:
+  std::string program_;
+  std::string params_;
+};
+
+/// CxxConfig is the config for the Full feature predictor.
+class LITE_API CxxConfig : public ConfigBase {
+  std::vector<Place> valid_places_;
+  std::string model_file_;
+  std::string param_file_;
+  std::shared_ptr<CxxModelBuffer> model_buffer_{nullptr};
+  std::vector<std::string> passes_internal_{};
+  bool quant_model_{false};  // Enable post_quant_dynamic in opt
+  QuantType quant_type_{QuantType::QUANT_INT16};
+  bool sparse_model_{false};  // Enable sparse_conv_detect_pass in opt
+  float sparse_threshold_{0.6f};
+  std::map<int, std::vector<std::shared_ptr<void>>>
+      preferred_inputs_for_warmup_;
+#ifdef LITE_WITH_CUDA
+  bool multi_stream_{false};
+#endif
+#ifdef LITE_WITH_MLU
+  lite_api::MLUCoreVersion mlu_core_version_{lite_api::MLUCoreVersion::MLU_270};
+  int mlu_core_number_{1};
+  DataLayoutType mlu_input_layout_{DATALAYOUT(kNCHW)};
+  std::vector<float> mlu_first_conv_mean_{};
+  std::vector<float> mlu_first_conv_std_{};
+#endif
+  // The custom configuration file or buffer for the NNAdapter subgraph
+  // partition, here is an example:
+  // op_type:in_var_name_0,in_var_name1:out_var_name_0,out_var_name1
+  // op_type::out_var_name_0
+  // op_type:in_var_name_0
+  // op_type
+  std::string nnadapter_subgraph_partition_config_path_;
+  std::string nnadapter_subgraph_partition_config_buffer_;
+  std::string mixed_precision_quantization_config_path_;
+  std::string mixed_precision_quantization_config_buffer_;
+
+ public:
+  void set_valid_places(const std::vector<Place>& x) { valid_places_ = x; }
+  void set_model_file(const std::string& path) { model_file_ = path; }
+  void set_param_file(const std::string& path) { param_file_ = path; }
+  void set_model_buffer(const char* model_buffer,
+                        size_t model_buffer_size,
+                        const char* param_buffer,
+                        size_t param_buffer_size) {
+    model_buffer_.reset(new CxxModelBuffer(
+        model_buffer, model_buffer_size, param_buffer, param_buffer_size));
+  }
+  void set_model_buffer(std::shared_ptr<CxxModelBuffer> model_buffer) {
+    model_buffer_ = model_buffer;
+  }
+  const CxxModelBuffer& get_model_buffer() const;
+  // internal inference to choose passes for model optimizing,
+  // it's designed for internal developer and not recommanded
+  // for comman users.
+  void set_passes_internal(
+      const std::vector<std::string>& passes_internal = {}) {
+    passes_internal_ = passes_internal;
+  }
+  const std::vector<std::string>& get_passes_internal() const {
+    return passes_internal_;
+  }
+  const std::vector<Place>& valid_places() const { return valid_places_; }
+  std::string model_file() const { return model_file_; }
+  std::string param_file() const { return param_file_; }
+  bool is_model_from_memory() const { return static_cast<bool>(model_buffer_); }
+  // note: `model_from_memory` has the same effect as `is_model_from_memory`,
+  // but is_model_from_memory is recommended and `model_from_memory` will be
+  // abandoned in v3.0.
+  bool model_from_memory() const { return static_cast<bool>(model_buffer_); }
+
+#ifdef LITE_WITH_CUDA
+  void set_multi_stream(bool multi_stream) { multi_stream_ = multi_stream; }
+  bool multi_stream() const { return multi_stream_; }
+#endif
+
+#ifdef LITE_WITH_MLU
+  // set MLU core version, which is used when compiling MLU kernels
+  void set_mlu_core_version(lite_api::MLUCoreVersion core_version);
+  // set MLU core number, which is used when compiling MLU kernels
+  void set_mlu_core_number(int core_number);
+  // whether use MLU's first conv kernel. First conv is a special kernel
+  // provided by MLU, its input is uint8, and also needs two 3-dimentional
+  // vectors which save all inputs' mean and std values
+  // set the 3-dimentional mean vector and 3-dimentional std vector used by
+  // MLU's first conv
+  void set_mlu_firstconv_param(const std::vector<float>& mean,
+                               const std::vector<float>& std);
+  // set MLU input layout. User can specify layout of input data to be NHWC,
+  // default is NCHW
+  void set_mlu_input_layout(DataLayoutType layout);
+
+  lite_api::MLUCoreVersion mlu_core_version() const;
+  int mlu_core_number() const;
+  DataLayoutType mlu_input_layout() const;
+  // std::pair<mean, std>
+  std::pair<std::vector<float>, std::vector<float>> mlu_firstconv_param() const;
+#endif
+
+  // XPU only, set the size of the workspace memory from L3 cache for the
+  // current thread.
+  // **DEPRECATED**, use set_xpu_l3_cache_method() in the future
+  void set_xpu_workspace_l3_size_per_thread(int l3_size = 0x4000000);
+  void set_xpu_l3_cache_method(size_t l3_size, bool locked = false);
+
+  void set_xpu_gm_workspace_method(size_t gm_size);
+
+  void set_xpu_conv_autotune(bool autotune = true,
+                             const std::string& autotune_file = "");
+
+  // XPU only, specify the target device ID for the current thread.
+  // **DEPRECATED**, use xpu_set_device() at the very beginning of each worker
+  // thread
+  void set_xpu_dev_per_thread(int dev_no = 0);
+
+  // XPU set multi_stream
+  void enable_xpu_multi_stream();
+
+  // **DEPRECATED**, use set_xpu_multi_encoder_method() in the future
+  void set_xpu_multi_encoder_precision(const std::string& precision = "int16");
+  void set_xpu_multi_encoder_method(const std::string& precision = "int16",
+                                    bool adaptive_seqlen = false);
+
+  // set input tensor for warmup.
+  // It is optional. If you set prefered_inputs, model wil run immediately when
+  // predictor is created
+  template <class T>
+  void set_preferred_inputs_for_warmup(const int group_idx,
+                                       const int tensor_idx,
+                                       const shape_t& shape,
+                                       const lod_t& lod = {},
+                                       const T fill_value = 0,
+                                       const void* data = nullptr);
+  const std::map<int, std::vector<std::shared_ptr<void>>>&
+  preferred_inputs_for_warmup() const {
+    return preferred_inputs_for_warmup_;
+  }
+
+  void set_quant_model(bool quant_model) { quant_model_ = quant_model; }
+  bool quant_model() const { return quant_model_; }
+  void set_quant_type(QuantType quant_type) { quant_type_ = quant_type; }
+  QuantType quant_type() const { return quant_type_; }
+
+  void set_sparse_model(bool sparse_model) { sparse_model_ = sparse_model; }
+  bool sparse_model() const { return sparse_model_; }
+  void set_sparse_threshold(float sparse_threshold) {
+    sparse_threshold_ = sparse_threshold;
+  }
+  float sparse_threshold() const { return sparse_threshold_; }
+
+  // Enable the custom subgraph partition for NNAdapter by providing the
+  // configuration file or buffer
+  void set_nnadapter_subgraph_partition_config_path(
+      const std::string& subgraph_partition_config_path) {
+    nnadapter_subgraph_partition_config_path_ = subgraph_partition_config_path;
+  }
+  const std::string& nnadapter_subgraph_partition_config_path() const {
+    return nnadapter_subgraph_partition_config_path_;
+  }
+  void set_nnadapter_subgraph_partition_config_buffer(
+      const std::string& subgraph_partition_config_buffer) {
+    nnadapter_subgraph_partition_config_buffer_ =
+        subgraph_partition_config_buffer;
+  }
+  const std::string& nnadapter_subgraph_partition_config_buffer() const {
+    return nnadapter_subgraph_partition_config_buffer_;
+  }
+  // Clear some ops' quant information to support mixed precision compute by
+  // configuration file or buffer
+  void set_nnadapter_mixed_precision_quantization_config_path(
+      const std::string& mixed_precision_quantization_config_path) {
+    mixed_precision_quantization_config_path_ =
+        mixed_precision_quantization_config_path;
+  }
+  const std::string& nnadapter_mixed_precision_quantization_config_path()
+      const {
+    return mixed_precision_quantization_config_path_;
+  }
+  void set_nnadapter_mixed_precision_quantization_config_buffer(
+      const std::string& mixed_precision_quantization_config_buffer) {
+    mixed_precision_quantization_config_buffer_ =
+        mixed_precision_quantization_config_buffer;
+  }
+  const std::string& nnadapter_mixed_precision_quantization_config_buffer()
+      const {
+    return mixed_precision_quantization_config_buffer_;
+  }
+};
+
+/// MobileConfig is the config for the light weight predictor, it will skip
+/// IR optimization or other unnecessary stages.
+class LITE_API MobileConfig : public ConfigBase {
+  // whether to load data from memory. Model data will be loaded from memory
+  // buffer if model_from_memory_ is true.
+  bool model_from_memory_{false};
+
+  // model data readed from file or memory buffer in combined format.
+  std::string lite_model_file_;
+
+  // NOTE: This is a deprecated variable and will be removed in latter release.
+  std::string model_buffer_;
+  std::string param_buffer_;
+
+ public:
+  // set model data in combined format, `set_model_from_file` refers to loading
+  // model from file, set_model_from_buffer refers to loading model from memory
+  // buffer
+  void set_model_from_file(const std::string& x);
+  void set_model_from_buffer(const std::string& x);
+  // return model data in lite_model_file_, which is in combined format.
+  const std::string& lite_model_file() const { return lite_model_file_; }
+
+  // return model_from_memory_, which indicates whether to load model from
+  // memory buffer.
+  bool is_model_from_memory() const { return model_from_memory_; }
+  // note: `model_from_memory` has the same effect as `is_model_from_memory`,
+  // but is_model_from_memory is recommended and `model_from_memory` will be
+  // abandoned in v3.0.
+  bool model_from_memory() const { return model_from_memory_; }
+
+  // NOTE: This is a deprecated API and will be removed in latter release.
+  void set_model_buffer(const char* model_buffer,
+                        size_t model_buffer_size,
+                        const char* param_buffer,
+                        size_t param_buffer_size);
+
+  // NOTE: This is a deprecated API and will be removed in latter release.
+  const std::string& model_buffer() const { return model_buffer_; }
+
+  // NOTE: This is a deprecated API and will be removed in latter release.
+  const std::string& param_buffer() const { return param_buffer_; }
+
+  // This is the method for allocating workspace_size according to L3Cache size
+  void SetArmL3CacheSize(
+      L3CacheSetMethod method = L3CacheSetMethod::kDeviceL3Cache,
+      int absolute_val = -1);
+};
+
+template <typename ConfigT>
+LITE_API std::shared_ptr<PaddlePredictor> CreatePaddlePredictor(const ConfigT&);
+
+}  // namespace lite_api
+}  // namespace paddle
+
+#endif  // NOLINT

libs/paddleocr/PaddleLite/cxx/include/paddle_image_preprocess.h

@@ -0,0 +1,274 @@
+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include <stdint.h>
+#include <stdio.h>
+#include <vector>
+#include "lite/api/paddle_api.h"
+#include "lite/api/paddle_place.h"
+
+namespace paddle {
+namespace lite {
+namespace utils {
+namespace cv {
+typedef paddle::lite_api::Tensor Tensor;
+typedef paddle::lite_api::DataLayoutType LayoutType;
+// color enum
+enum ImageFormat {
+  RGBA = 0,
+  BGRA,
+  RGB,
+  BGR,
+  GRAY,
+  NV21 = 11,
+  NV12,
+  YUV420SP,
+  YUV420P,
+  YUV422,
+  YUV444
+};
+// flip enum
+enum FlipParam {
+  XY = -1,  // flip along the XY axis
+  X = 0,    // flip along the X axis
+  Y         // flip along the Y axis
+};
+// transform param
+typedef struct {
+  int ih;                // input height
+  int iw;                // input width
+  int oh;                // outpu theight
+  int ow;                // output width
+  FlipParam flip_param;  // flip, support x, y, xy
+  float rotate_param;    // rotate, support 90, 180, 270
+} TransParam;
+
+class ImagePreprocess {
+ public:
+  /*
+  * init
+  * param srcFormat: input image color
+  * param dstFormat: output image color
+  * param param: input image parameter, egs: input size
+  */
+  ImagePreprocess(ImageFormat srcFormat,
+                  ImageFormat dstFormat,
+                  TransParam param);
+
+  /*
+  * image color convert
+  * support NV12/NV21_to_BGR(RGB), NV12/NV21_to_BGRA(RGBA),
+  * BGR(RGB)and BGRA(RGBA) transform,
+  * BGR(RGB)and RGB(BGR) transform,
+  * BGR(RGB)and RGBA(BGRA) transform,
+  * BGR(RGB) and GRAY transform,
+  * BGRA(RGBA) and GRAY transform,
+  * param src: input image data
+  * param dst: output image data
+  */
+  void image_convert(const uint8_t* src, uint8_t* dst);
+
+  /*
+  * image color convert
+  * support NV12/NV21_to_BGR(RGB), NV12/NV21_to_BGRA(RGBA),
+  * BGR(RGB)and BGRA(RGBA) transform,
+  * BGR(RGB)and RGB(BGR) transform,
+  * BGR(RGB)and RGBA(BGRA) transform,
+  * BGR(RGB)and GRAY transform,
+  * BGRA(RGBA) and GRAY transform,
+  * param src: input image data
+  * param dst: output image data
+  * param srcFormat: input image image format support: GRAY, NV12(NV21),
+  * BGR(RGB) and BGRA(RGBA)
+  * param dstFormat: output image image format, support GRAY, BGR(RGB) and
+  * BGRA(RGBA)
+  */
+  void image_convert(const uint8_t* src,
+                     uint8_t* dst,
+                     ImageFormat srcFormat,
+                     ImageFormat dstFormat);
+
+  /*
+  * image color convert
+  * support NV12/NV21_to_BGR(RGB), NV12/NV21_to_BGRA(RGBA),
+  * BGR(RGB)and BGRA(RGBA) transform,
+  * BGR(RGB)and RGB(BGR) transform,
+  * BGR(RGB)and RGBA(BGRA) transform,
+  * BGR(RGB)and GRAY transform,
+  * BGRA(RGBA) and GRAY transform,
+  * param src: input image data
+  * param dst: output image data
+  * param srcFormat: input image image format support: GRAY, NV12(NV21),
+  * BGR(RGB) and BGRA(RGBA)
+  * param dstFormat: output image image format, support GRAY, BGR(RGB) and
+  * BGRA(RGBA)
+  * param srcw: input image width
+  * param srch: input image height
+  */
+  void image_convert(const uint8_t* src,
+                     uint8_t* dst,
+                     ImageFormat srcFormat,
+                     ImageFormat dstFormat,
+                     int srcw,
+                     int srch);
+
+  /*
+  * image resize, use bilinear method
+  * support image format: 1-channel image (egs: GRAY, 2-channel image (egs:
+  * NV12, NV21), 3-channel(egs: BGR), 4-channel(egs: BGRA)
+  * param src: input image data
+  * param dst: output image data
+  */
+  void image_resize(const uint8_t* src, uint8_t* dst);
+
+  /*
+   image resize, use bilinear method
+  * support image format: 1-channel image (egs: GRAY, 2-channel image (egs:
+  NV12, NV21), 3-channel image(egs: BGR), 4-channel image(egs: BGRA)
+  * param src: input image data
+  * param dst: output image data
+  * param srcw: input image width
+  * param srch: input image height
+  * param dstw: output image width
+  * param dsth: output image height
+  */
+  void image_resize(const uint8_t* src,
+                    uint8_t* dst,
+                    ImageFormat srcFormat,
+                    int srcw,
+                    int srch,
+                    int dstw,
+                    int dsth);
+
+  /*
+  * image Rotate
+  * support 90, 180 and 270 Rotate process
+  * color format support 1-channel image, 3-channel image and 4-channel image
+  * param src: input image data
+  * param dst: output image data
+  */
+  void image_rotate(const uint8_t* src, uint8_t* dst);
+
+  /*
+  * image Rotate
+  * support 90, 180 and 270 Rotate process
+  * color format support 1-channel image, 3-channel image and 4-channel image
+  * param src: input image data
+  * param dst: output image data
+  * param srcFormat: input image format, support GRAY, BGR(RGB) and BGRA(RGBA)
+  * param srcw: input image width
+  * param srch: input image height
+  * param degree: Rotate degree, support 90, 180 and 270
+  */
+  void image_rotate(const uint8_t* src,
+                    uint8_t* dst,
+                    ImageFormat srcFormat,
+                    int srcw,
+                    int srch,
+                    float degree);
+
+  /*
+  * image Flip
+  * support X, Y and XY flip process
+  * color format support 1-channel image, 3-channel image and 4-channel image
+  * param src: input image data
+  * param dst: output image data
+  */
+  void image_flip(const uint8_t* src, uint8_t* dst);
+
+  /*
+  * image Flip
+  * support X, Y and XY flip process
+  * color format support 1-channel image, 3-channel image and 4-channel image
+  * param src: input image data
+  * param dst: output image data
+  * param srcFormat: input image format, support GRAY, BGR(RGB) and BGRA(RGBA)
+  * param srcw: input image width
+  * param srch: input image height
+  * param flip_param: flip parameter, support X, Y and XY
+  */
+  void image_flip(const uint8_t* src,
+                  uint8_t* dst,
+                  ImageFormat srcFormat,
+                  int srcw,
+                  int srch,
+                  FlipParam flip_param);
+
+  /*
+  * change image data to tensor data
+  * support image format is GRAY, BGR(RGB) and BGRA(RGBA), Data layout is NHWC
+  * and
+  * NCHW
+  * param src: input image data
+  * param dstTensor: output tensor data
+  * param layout: output tensor layout，support NHWC and NCHW
+  * param means: means of image
+  * param scales: scales of image
+  */
+  void image_to_tensor(const uint8_t* src,
+                       Tensor* dstTensor,
+                       LayoutType layout,
+                       float* means,
+                       float* scales);
+
+  /*
+  * change image data to tensor data
+  * support image format is GRAY, BGR(RGB) and BGRA(RGBA), Data layout is NHWC
+  * and
+  * NCHW
+  * param src: input image data
+  * param dstTensor: output tensor data
+  * param srcFormat: input image format, support BGR(RGB) and BGRA(RGBA)
+  * param srcw: input image width
+  * param srch: input image height
+  * param layout: output tensor layout，support NHWC and NCHW
+  * param means: means of image
+  * param scales: scales of image
+  */
+  void image_to_tensor(const uint8_t* src,
+                       Tensor* dstTensor,
+                       ImageFormat srcFormat,
+                       int srcw,
+                       int srch,
+                       LayoutType layout,
+                       float* means,
+                       float* scales);
+
+  /*
+  * image crop process
+  * color format support 1-channel image, 3-channel image and 4-channel image
+  * param src: input image data
+  * param dst: output image data
+  */
+  void image_crop(const uint8_t* src,
+                  uint8_t* dst,
+                  ImageFormat srcFormat,
+                  int srcw,
+                  int srch,
+                  int left_x,
+                  int left_y,
+                  int dstw,
+                  int dsth);
+
+ private:
+  ImageFormat srcFormat_;
+  ImageFormat dstFormat_;
+  TransParam transParam_;
+};
+}  // namespace cv
+}  // namespace utils
+}  // namespace lite
+}  // namespace paddle

libs/paddleocr/PaddleLite/cxx/include/paddle_lite_factory_helper.h

@@ -0,0 +1,44 @@
+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+/*
+ * This file defines some MACROS that explicitly determine the op, kernel, mir
+ * passes used in the inference lib.
+ */
+#pragma once
+
+// some platform-independent defintion
+
+#if defined(_WIN32)
+#define UNUSED
+#define __builtin_expect(EXP, C) (EXP)
+#else
+#define UNUSED __attribute__((unused))
+#endif
+
+#define USE_LITE_OP(op_type__)       \
+  extern int touch_op_##op_type__(); \
+  int LITE_OP_REGISTER_FAKE(op_type__) UNUSED = touch_op_##op_type__();
+
+#define USE_LITE_KERNEL(op_type__, target__, precision__, layout__, alias__) \
+  extern int touch_##op_type__##target__##precision__##layout__##alias__();  \
+  int op_type__##target__##precision__##layout__##alias__##__use_lite_kernel \
+      UNUSED = touch_##op_type__##target__##precision__##layout__##alias__();
+
+#define USE_MIR_PASS(name__)                      \
+  extern bool mir_pass_registry##name__##_fake(); \
+  static bool mir_pass_usage##name__ UNUSED =     \
+      mir_pass_registry##name__##_fake();
+
+#define LITE_OP_REGISTER_FAKE(op_type__) op_type__##__registry__

libs/paddleocr/PaddleLite/cxx/include/paddle_place.h

@@ -0,0 +1,278 @@
+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+#include <set>
+#include <string>
+
+// Generic helper definitions for shared library support
+#if defined _WIN32 || defined __CYGWIN__
+#define PADDLE_LITE_HELPER_DLL_IMPORT __declspec(dllimport)
+#define PADDLE_LITE_HELPER_DLL_EXPORT __declspec(dllexport)
+#define PADDLE_LITE_HELPER_DLL_LOCAL
+#else
+#if __GNUC__ >= 4
+#define PADDLE_LITE_HELPER_DLL_IMPORT __attribute__((visibility("default")))
+#define PADDLE_LITE_HELPER_DLL_EXPORT __attribute__((visibility("default")))
+#else
+#define PADDLE_LITE_HELPER_DLL_IMPORT
+#define PADDLE_LITE_HELPER_DLL_EXPORT
+#endif
+#endif
+
+#ifdef LITE_ON_TINY_PUBLISH
+#define LITE_API PADDLE_LITE_HELPER_DLL_EXPORT
+#define LITE_API_IMPORT PADDLE_LITE_HELPER_DLL_IMPORT
+#else
+#define LITE_API
+#define LITE_API_IMPORT
+#endif
+
+namespace paddle {
+namespace lite_api {
+
+enum class TargetType : int {
+  kUnk = 0,
+  kHost = 1,
+  kX86 = 2,
+  kCUDA = 3,
+  kARM = 4,
+  kOpenCL = 5,
+  kAny = 6,  // any target
+  kFPGA = 7,
+  kNPU = 8,
+  kXPU = 9,
+  kBM = 10,
+  kMLU = 11,
+  kRKNPU = 12,
+  kAPU = 13,
+  kHuaweiAscendNPU = 14,
+  kImaginationNNA = 15,
+  kIntelFPGA = 16,
+  kMetal = 17,
+  kNNAdapter = 18,
+  NUM = 19,  // number of fields.
+};
+enum class PrecisionType : int {
+  kUnk = 0,
+  kFloat = 1,
+  kInt8 = 2,
+  kInt32 = 3,
+  kAny = 4,  // any precision
+  kFP16 = 5,
+  kBool = 6,
+  kInt64 = 7,
+  kInt16 = 8,
+  kUInt8 = 9,
+  kFP64 = 10,
+  NUM = 11,  // number of fields.
+};
+enum class DataLayoutType : int {
+  kUnk = 0,
+  kNCHW = 1,
+  kNHWC = 3,
+  kImageDefault = 4,  // for opencl image2d
+  kImageFolder = 5,   // for opencl image2d
+  kImageNW = 6,       // for opencl image2d
+  kAny = 2,           // any data layout
+  kMetalTexture2DArray = 7,
+  kMetalTexture2D = 8,
+  NUM = 9,  // number of fields.
+};
+
+typedef enum {
+  LITE_POWER_HIGH = 0,
+  LITE_POWER_LOW = 1,
+  LITE_POWER_FULL = 2,
+  LITE_POWER_NO_BIND = 3,
+  LITE_POWER_RAND_HIGH = 4,
+  LITE_POWER_RAND_LOW = 5
+} PowerMode;
+
+typedef enum {
+  CL_TUNE_NONE = 0,
+  CL_TUNE_RAPID = 1,
+  CL_TUNE_NORMAL = 2,
+  CL_TUNE_EXHAUSTIVE = 3
+} CLTuneMode;
+
+typedef enum {
+  CL_PRECISION_AUTO = 0,
+  CL_PRECISION_FP32 = 1,
+  CL_PRECISION_FP16 = 2
+} CLPrecisionType;
+
+typedef enum { MLU_220 = 0, MLU_270 = 1 } MLUCoreVersion;
+
+enum class ActivationType : int {
+  kIndentity = 0,
+  kRelu = 1,
+  kRelu6 = 2,
+  kPRelu = 3,
+  kLeakyRelu = 4,
+  kSigmoid = 5,
+  kTanh = 6,
+  kSwish = 7,
+  kExp = 8,
+  kAbs = 9,
+  kHardSwish = 10,
+  kReciprocal = 11,
+  kThresholdedRelu = 12,
+  kElu = 13,
+  kHardSigmoid = 14,
+  kLog = 15,
+  kSigmoid_v2 = 16,
+  kTanh_v2 = 17,
+  kGelu = 18,
+  kErf = 19,
+  kSign = 20,
+  kSoftPlus = 21,
+  kMish = 22,
+  NUM = 23,
+};
+
+static size_t PrecisionTypeLength(PrecisionType type) {
+  switch (type) {
+    case PrecisionType::kFloat:
+      return 4;
+    case PrecisionType::kFP64:
+      return 8;
+    case PrecisionType::kUInt8:
+      return 1;
+    case PrecisionType::kInt8:
+      return 1;
+    case PrecisionType::kInt32:
+      return 4;
+    case PrecisionType::kInt64:
+      return 8;
+    case PrecisionType::kFP16:
+      return 2;
+    case PrecisionType::kInt16:
+      return 2;
+    default:
+      return 0;
+  }
+}
+
+enum class QuantType : int {
+  QUANT_INT8,
+  QUANT_INT16,
+};
+
+template <typename T>
+struct PrecisionTypeTrait {
+  constexpr static PrecisionType Type() { return PrecisionType::kUnk; }
+};
+
+#define _ForEachPrecisionTypeHelper(callback, cpp_type, precision_type) \
+  callback(cpp_type, ::paddle::lite_api::PrecisionType::precision_type);
+
+#define _ForEachPrecisionType(callback)                   \
+  _ForEachPrecisionTypeHelper(callback, bool, kBool);     \
+  _ForEachPrecisionTypeHelper(callback, float, kFloat);   \
+  _ForEachPrecisionTypeHelper(callback, double, kFP64);   \
+  _ForEachPrecisionTypeHelper(callback, uint8_t, kUInt8); \
+  _ForEachPrecisionTypeHelper(callback, int8_t, kInt8);   \
+  _ForEachPrecisionTypeHelper(callback, int16_t, kInt16); \
+  _ForEachPrecisionTypeHelper(callback, int, kInt32);     \
+  _ForEachPrecisionTypeHelper(callback, int64_t, kInt64);
+
+#define DefinePrecisionTypeTrait(cpp_type, precision_type)           \
+  template <>                                                        \
+  struct PrecisionTypeTrait<cpp_type> {                              \
+    constexpr static PrecisionType Type() { return precision_type; } \
+  }
+
+_ForEachPrecisionType(DefinePrecisionTypeTrait);
+
+#ifdef ENABLE_ARM_FP16
+typedef __fp16 float16_t;
+_ForEachPrecisionTypeHelper(DefinePrecisionTypeTrait, float16_t, kFP16);
+#endif
+
+#undef _ForEachPrecisionTypeHelper
+#undef _ForEachPrecisionType
+#undef DefinePrecisionTypeTrait
+
+#define TARGET(item__) paddle::lite_api::TargetType::item__
+#define PRECISION(item__) paddle::lite_api::PrecisionType::item__
+#define DATALAYOUT(item__) paddle::lite_api::DataLayoutType::item__
+
+const std::string& ActivationTypeToStr(ActivationType act);
+
+const std::string& TargetToStr(TargetType target);
+
+const std::string& PrecisionToStr(PrecisionType precision);
+
+const std::string& DataLayoutToStr(DataLayoutType layout);
+
+const std::string& TargetRepr(TargetType target);
+
+const std::string& PrecisionRepr(PrecisionType precision);
+
+const std::string& DataLayoutRepr(DataLayoutType layout);
+
+const std::string& CLTuneModeToStr(CLTuneMode mode);
+
+const std::string& CLPrecisionTypeToStr(CLPrecisionType type);
+
+// Get a set of all the elements represented by the target.
+std::set<TargetType> ExpandValidTargets(TargetType target = TARGET(kAny));
+
+// Get a set of all the elements represented by the precision.
+std::set<PrecisionType> ExpandValidPrecisions(
+    PrecisionType precision = PRECISION(kAny));
+
+// Get a set of all the elements represented by the layout.
+std::set<DataLayoutType> ExpandValidLayouts(
+    DataLayoutType layout = DATALAYOUT(kAny));
+
+/*
+ * Place specifies the execution context of a Kernel or input/output for a
+ * kernel. It is used to make the analysis of the MIR more clear and accurate.
+ */
+struct LITE_API Place {
+  TargetType target{TARGET(kUnk)};
+  PrecisionType precision{PRECISION(kUnk)};
+  DataLayoutType layout{DATALAYOUT(kUnk)};
+  int16_t device{0};  // device ID
+
+  Place() = default;
+  Place(TargetType target,
+        PrecisionType precision = PRECISION(kFloat),
+        DataLayoutType layout = DATALAYOUT(kNCHW),
+        int16_t device = 0)
+      : target(target), precision(precision), layout(layout), device(device) {}
+
+  bool is_valid() const {
+    return target != TARGET(kUnk) && precision != PRECISION(kUnk) &&
+           layout != DATALAYOUT(kUnk);
+  }
+
+  size_t hash() const;
+
+  bool operator==(const Place& other) const {
+    return target == other.target && precision == other.precision &&
+           layout == other.layout && device == other.device;
+  }
+
+  bool operator!=(const Place& other) const { return !(*this == other); }
+
+  friend bool operator<(const Place& a, const Place& b);
+
+  std::string DebugString() const;
+};
+
+}  // namespace lite_api
+}  // namespace paddle

libs/paddleocr/PaddleLite/cxx/include/paddle_use_kernels.h

@@ -0,0 +1,408 @@
+#pragma once
+#include "paddle_lite_factory_helper.h"
+
+USE_LITE_KERNEL(gather_nd, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(relu_clipped, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(swish, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(log, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(exp, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(floor, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(hard_sigmoid, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sqrt, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(rsqrt, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(square, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(hard_swish, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(reciprocal, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(abs, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(gelu, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(erf, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sign, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(softplus, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(mish, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(pow, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(where, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(assign_value, kHost, kAny, kNCHW, def);
+USE_LITE_KERNEL(tril_triu, kHost, kAny, kNCHW, float32);
+USE_LITE_KERNEL(lstm, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(lstm, kARM, kInt8, kNCHW, def);
+USE_LITE_KERNEL(split, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(split, kHost, kFloat, kNCHW, int32);
+USE_LITE_KERNEL(split, kHost, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(split, kHost, kInt64, kNCHW, def);
+USE_LITE_KERNEL(gaussian_random, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(feed, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(cos, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(conditional_block, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(read_from_array, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(deformable_conv, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(distribute_fpn_proposals, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(roi_align, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(one_hot, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(one_hot_v2, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(one_hot_v2, kHost, kAny, kAny, one_hot_v2_int32);
+USE_LITE_KERNEL(unique_with_counts, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(arg_max, kARM, kAny, kNCHW, fp32);
+USE_LITE_KERNEL(arg_max, kARM, kAny, kNCHW, int64);
+USE_LITE_KERNEL(arg_max, kARM, kAny, kNCHW, int32);
+USE_LITE_KERNEL(arg_max, kARM, kAny, kNCHW, int16);
+USE_LITE_KERNEL(arg_max, kARM, kAny, kNCHW, uint8);
+USE_LITE_KERNEL(lod_array_length, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(flip, kHost, kAny, kNCHW, flip_fp32);
+USE_LITE_KERNEL(flip, kHost, kAny, kNCHW, flip_i64);
+USE_LITE_KERNEL(reduce_all, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(reduce_any, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(stack, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(stack, kHost, kFloat, kAny, int32_def);
+USE_LITE_KERNEL(stack, kHost, kFloat, kAny, int64_def);
+USE_LITE_KERNEL(elementwise_add, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_add, kARM, kInt32, kNCHW, def);
+USE_LITE_KERNEL(elementwise_add, kARM, kInt64, kNCHW, def);
+USE_LITE_KERNEL(elementwise_add, kARM, kFloat, kNCHW, int32);
+USE_LITE_KERNEL(elementwise_add, kARM, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(fusion_elementwise_add_activation, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_sub, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_sub, kARM, kInt32, kNCHW, def);
+USE_LITE_KERNEL(elementwise_sub, kARM, kFloat, kNCHW, int32);
+USE_LITE_KERNEL(elementwise_sub, kARM, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(fusion_elementwise_sub_activation, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_mul, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_mul, kARM, kInt32, kNCHW, def);
+USE_LITE_KERNEL(elementwise_mul, kARM, kInt64, kNCHW, def);
+USE_LITE_KERNEL(elementwise_mul, kARM, kFloat, kNCHW, int32);
+USE_LITE_KERNEL(elementwise_mul, kARM, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(fusion_elementwise_mul_activation, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(fusion_elementwise_mul_activation, kARM, kInt64, kNCHW, def);
+USE_LITE_KERNEL(elementwise_max, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(fusion_elementwise_max_activation, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_min, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(fusion_elementwise_min_activation, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_div, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_div, kARM, kInt32, kNCHW, def);
+USE_LITE_KERNEL(elementwise_div, kARM, kInt64, kNCHW, def);
+USE_LITE_KERNEL(fusion_elementwise_div_activation, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_mod, kARM, kInt64, kNCHW, def);
+USE_LITE_KERNEL(elementwise_mod, kARM, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(elementwise_mod, kARM, kFloat, kNCHW, int32_mod);
+USE_LITE_KERNEL(elementwise_pow, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_pow, kARM, kInt32, kNCHW, def);
+USE_LITE_KERNEL(elementwise_floordiv, kARM, kInt32, kNCHW, def);
+USE_LITE_KERNEL(elementwise_floordiv, kARM, kInt64, kNCHW, def);
+USE_LITE_KERNEL(elementwise_floordiv, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elementwise_floordiv, kARM, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(softmax, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(prior_box, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(gru, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(gru, kARM, kInt8, kNCHW, def);
+USE_LITE_KERNEL(is_empty, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(unsqueeze, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(unsqueeze2, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(sequence_expand_as, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sequence_expand_as, kARM, kFloat, kNCHW, int32);
+USE_LITE_KERNEL(sequence_expand_as, kARM, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(fill_constant_batch_size_like, kHost, kAny, kNCHW, def);
+USE_LITE_KERNEL(fill_zeros_like, kHost, kFloat, kNCHW, float32);
+USE_LITE_KERNEL(fill_zeros_like, kHost, kFloat, kNCHW, int32);
+USE_LITE_KERNEL(fill_zeros_like, kHost, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(sum, kARM, kFloat, kNCHW, sum_i32);
+USE_LITE_KERNEL(sum, kARM, kFloat, kNCHW, sum_i64);
+USE_LITE_KERNEL(sum, kARM, kFloat, kNCHW, sum_fp32);
+USE_LITE_KERNEL(relu, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(leaky_relu, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(relu_clipped, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(prelu, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sigmoid, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(tanh, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(swish, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(relu6, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(log, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(exp, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(floor, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(hard_sigmoid, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(rsqrt, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(square, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(hard_swish, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(reciprocal, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(abs, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(thresholded_relu, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elu, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(softplus, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(reduce_mean, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(range, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(range, kHost, kInt64, kAny, def);
+USE_LITE_KERNEL(range, kHost, kInt32, kAny, def);
+USE_LITE_KERNEL(range, kHost, kFloat, kAny, int32);
+USE_LITE_KERNEL(range, kHost, kFloat, kAny, int64);
+USE_LITE_KERNEL(beam_search_decode, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(squeeze, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(squeeze2, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(group_norm, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sequence_conv, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(write_to_array, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(unstack, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(unstack, kHost, kFloat, kAny, unstack_int32);
+USE_LITE_KERNEL(generate_proposals_v2, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(layer_norm, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(crop_tensor, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(crop_tensor, kHost, kFloat, kAny, int32_precision);
+USE_LITE_KERNEL(sequence_mask, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sequence_mask, kHost, kFloat, kNCHW, int32);
+USE_LITE_KERNEL(sequence_mask, kHost, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(matrix_nms, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(reverse, kHost, kAny, kNCHW, fp32);
+USE_LITE_KERNEL(sequence_expand, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sequence_expand, kHost, kFloat, kNCHW, int32);
+USE_LITE_KERNEL(sequence_expand, kHost, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(reduce_prod, kARM, kInt32, kNCHW, def);
+USE_LITE_KERNEL(reduce_prod, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(reduce_prod, kARM, kFloat, kNCHW, reduce_prod_i64);
+USE_LITE_KERNEL(reduce_prod, kARM, kFloat, kNCHW, int32);
+USE_LITE_KERNEL(pad3d, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(cos_sim, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(relu, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(leaky_relu, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(prelu, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sigmoid, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(tanh, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(relu6, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(thresholded_relu, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(elu, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(polygon_box_transform, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(unfold, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(unfold, kHost, kFloat, kNCHW, def_int32);
+USE_LITE_KERNEL(unfold, kHost, kFloat, kNCHW, def_int64);
+USE_LITE_KERNEL(unfold, kHost, kInt8, kNCHW, def_int8);
+USE_LITE_KERNEL(pixel_shuffle, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(increment, kHost, kAny, kNCHW, def);
+USE_LITE_KERNEL(ctc_align, kHost, kInt64, kNCHW, def);
+USE_LITE_KERNEL(ctc_align, kHost, kInt32, kNCHW, def);
+USE_LITE_KERNEL(instance_norm, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(linspace, kHost, kFloat, kAny, float32);
+USE_LITE_KERNEL(linspace, kHost, kInt32, kAny, int32);
+USE_LITE_KERNEL(print, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(expand, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(assign, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(assign, kHost, kAny, kAny, def_tensor_array);
+USE_LITE_KERNEL(generate_proposals, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(lrn, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(cast, kHost, kAny, kNCHW, def);
+USE_LITE_KERNEL(log_softmax, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(beam_search, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(concat, kARM, kAny, kNCHW, def);
+USE_LITE_KERNEL(scatter, kARM, kFloat, kNCHW, ids_int64);
+USE_LITE_KERNEL(scatter, kARM, kFloat, kNCHW, ids_int32);
+USE_LITE_KERNEL(shuffle_channel, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(meshgrid, kHost, kFloat, kAny, float32);
+USE_LITE_KERNEL(meshgrid, kHost, kFloat, kAny, int32);
+USE_LITE_KERNEL(split_lod_tensor, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(pool2d, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(scale, kARM, kFloat, kNCHW, int32);
+USE_LITE_KERNEL(scale, kARM, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(scale, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(scale, kARM, kInt32, kNCHW, def);
+USE_LITE_KERNEL(scale, kARM, kInt64, kNCHW, def);
+USE_LITE_KERNEL(transpose, kARM, kAny, kNCHW, def);
+USE_LITE_KERNEL(transpose2, kARM, kAny, kNCHW, def);
+USE_LITE_KERNEL(depthwise_conv2d_transpose, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(depthwise_conv2d_transpose, kARM, kInt8, kNCHW, fp32_out);
+USE_LITE_KERNEL(depthwise_conv2d_transpose, kARM, kInt8, kNCHW, int8_out);
+USE_LITE_KERNEL(grid_sampler, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(axpy, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(tensor_array_to_tensor, kHost, kAny, kNCHW, def);
+USE_LITE_KERNEL(merge_lod_tensor, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(reshape, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(reshape2, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(flatten, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(flatten2, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(sparse_conv2d, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sparse_conv2d, kARM, kInt8, kNCHW, int8_fp32_out);
+USE_LITE_KERNEL(sparse_conv2d, kARM, kInt8, kNCHW, int8_int8_out);
+USE_LITE_KERNEL(lod_reset, kHost, kAny, kNCHW, def);
+USE_LITE_KERNEL(expand_as, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(expand_as, kHost, kFloat, kAny, int64);
+USE_LITE_KERNEL(calib, kARM, kInt8, kNCHW, fp32_to_int8);
+USE_LITE_KERNEL(calib, kARM, kInt32, kNCHW, int32_to_fp32);
+USE_LITE_KERNEL(calib, kARM, kInt32, kNCHW, int32_to_int64);
+USE_LITE_KERNEL(calib, kARM, kInt32, kNCHW, fp32_to_int32);
+USE_LITE_KERNEL(calib, kARM, kInt64, kNCHW, int64_to_fp32);
+USE_LITE_KERNEL(calib, kARM, kInt64, kNCHW, fp32_to_int64);
+USE_LITE_KERNEL(calib, kARM, kInt8, kNCHW, int8_to_fp32);
+USE_LITE_KERNEL(calib, kARM, kInt64, kNCHW, int64_to_int32);
+USE_LITE_KERNEL(calib_once, kARM, kInt8, kNCHW, fp32_to_int8);
+USE_LITE_KERNEL(calib_once, kARM, kInt8, kNCHW, int8_to_fp32);
+USE_LITE_KERNEL(calib_once, kARM, kInt64, kNCHW, int64_to_int32);
+USE_LITE_KERNEL(clip, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(select_input, kHost, kAny, kNCHW, def);
+USE_LITE_KERNEL(reduce_min, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(reduce_min, kARM, kFloat, kNCHW, def_int64);
+USE_LITE_KERNEL(sequence_pad, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sequence_pad, kHost, kFloat, kNCHW, int32);
+USE_LITE_KERNEL(sequence_pad, kHost, kFloat, kNCHW, int64);
+USE_LITE_KERNEL(logical_xor, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(logical_and, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(logical_or, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(logical_not, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(fill_constant, kHost, kAny, kNCHW, def);
+USE_LITE_KERNEL(yolo_box, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(mean, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(collect_fpn_proposals, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(matmul_v2, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(matmul_v2, kARM, kInt8, kNCHW, def);
+USE_LITE_KERNEL(inverse, kHost, kFloat, kNCHW, fp32);
+USE_LITE_KERNEL(top_k, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(uniform_random, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(matmul, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(matmul, kARM, kInt8, kNCHW, def);
+USE_LITE_KERNEL(norm, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(p_norm, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sequence_unpad, kHost, kFloat, kAny, float32);
+USE_LITE_KERNEL(sequence_unpad, kHost, kFloat, kAny, int64);
+USE_LITE_KERNEL(where_index, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(layout, kARM, kFloat, kNCHW, nchw2nhwc);
+USE_LITE_KERNEL(layout, kARM, kFloat, kNCHW, nhwc2nchw);
+USE_LITE_KERNEL(layout, kARM, kInt8, kNCHW, int8_nchw2nhwc);
+USE_LITE_KERNEL(layout, kARM, kInt8, kNCHW, int8_nhwc2nchw);
+USE_LITE_KERNEL(layout_once, kARM, kFloat, kNCHW, nchw2nhwc);
+USE_LITE_KERNEL(layout_once, kARM, kFloat, kNCHW, nhwc2nchw);
+USE_LITE_KERNEL(layout_once, kARM, kInt8, kNCHW, int8_nchw2nhwc);
+USE_LITE_KERNEL(layout_once, kARM, kInt8, kNCHW, int8_nhwc2nchw);
+USE_LITE_KERNEL(crop, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(crop, kHost, kInt32, kAny, def);
+USE_LITE_KERNEL(pad2d, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(argsort, kHost, kFloat, kAny, argsort_fp32);
+USE_LITE_KERNEL(argsort, kHost, kFloat, kAny, argsort_int32);
+USE_LITE_KERNEL(argsort, kHost, kFloat, kAny, argsort_int64);
+USE_LITE_KERNEL(equal, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(equal, kHost, kInt64, kAny, def);
+USE_LITE_KERNEL(equal, kHost, kFloat, kAny, int64);
+USE_LITE_KERNEL(equal, kHost, kInt32, kAny, def);
+USE_LITE_KERNEL(equal, kHost, kFloat, kAny, int32);
+USE_LITE_KERNEL(not_equal, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(not_equal, kHost, kFloat, kAny, int32);
+USE_LITE_KERNEL(not_equal, kHost, kFloat, kAny, int64);
+USE_LITE_KERNEL(less_than, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(less_than, kHost, kInt32, kAny, def);
+USE_LITE_KERNEL(less_than, kHost, kFloat, kAny, int32);
+USE_LITE_KERNEL(less_than, kHost, kInt64, kAny, def);
+USE_LITE_KERNEL(less_than, kHost, kFloat, kAny, int64);
+USE_LITE_KERNEL(less_equal, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(less_equal, kHost, kInt64, kAny, def);
+USE_LITE_KERNEL(less_equal, kHost, kFloat, kAny, int64);
+USE_LITE_KERNEL(less_equal, kHost, kFloat, kAny, int32);
+USE_LITE_KERNEL(greater_than, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(greater_than, kHost, kFloat, kAny, def_bool);
+USE_LITE_KERNEL(greater_than, kHost, kFloat, kAny, def_int32);
+USE_LITE_KERNEL(greater_than, kHost, kInt64, kAny, def);
+USE_LITE_KERNEL(greater_than, kHost, kFloat, kAny, def_int64);
+USE_LITE_KERNEL(greater_equal, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(greater_equal, kHost, kFloat, kAny, def_int64);
+USE_LITE_KERNEL(greater_equal, kHost, kFloat, kAny, def_int32);
+USE_LITE_KERNEL(sampling_id, kHost, kAny, kAny, float32);
+USE_LITE_KERNEL(pixel_shuffle, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(arg_max, kHost, kAny, kNCHW, fp32);
+USE_LITE_KERNEL(arg_max, kHost, kAny, kNCHW, int64);
+USE_LITE_KERNEL(arg_max, kHost, kAny, kNCHW, int32);
+USE_LITE_KERNEL(arg_max, kHost, kAny, kNCHW, int16);
+USE_LITE_KERNEL(arg_max, kHost, kAny, kNCHW, uint8);
+USE_LITE_KERNEL(anchor_generator, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(write_back, kHost, kAny, kAny, write_back);
+USE_LITE_KERNEL(correlation, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(im2sequence, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(batch_norm, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sync_batch_norm, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(gather_tree, kHost, kFloat, kAny, int32);
+USE_LITE_KERNEL(gather_tree, kHost, kFloat, kAny, int64);
+USE_LITE_KERNEL(flatten_contiguous_range, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(lookup_table, kARM, kAny, kNCHW, def);
+USE_LITE_KERNEL(lookup_table_v2, kARM, kAny, kNCHW, def);
+USE_LITE_KERNEL(top_k_v2, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(box_clip, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(box_coder, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(conv2d_transpose, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(conv2d_transpose, kARM, kInt8, kNCHW, fp32_out);
+USE_LITE_KERNEL(conv2d_transpose, kARM, kInt8, kNCHW, int8_out);
+USE_LITE_KERNEL(unbind, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(unbind, kHost, kFloat, kNCHW, def_int64);
+USE_LITE_KERNEL(shape, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(bilinear_interp, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(nearest_interp, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(bilinear_interp_v2, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(nearest_interp_v2, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(conv2d, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(depthwise_conv2d, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(conv2d, kARM, kInt8, kNCHW, int8_out);
+USE_LITE_KERNEL(conv2d, kARM, kInt8, kNCHW, fp32_out);
+USE_LITE_KERNEL(depthwise_conv2d, kARM, kInt8, kNCHW, int8_out);
+USE_LITE_KERNEL(depthwise_conv2d, kARM, kInt8, kNCHW, fp32_out);
+USE_LITE_KERNEL(reduce_max, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(reduce_max, kARM, kFloat, kNCHW, i64);
+USE_LITE_KERNEL(index_select, kHost, kAny, kNCHW, fp32);
+USE_LITE_KERNEL(index_select, kHost, kAny, kNCHW, int32);
+USE_LITE_KERNEL(index_select, kHost, kAny, kNCHW, int16);
+USE_LITE_KERNEL(index_select, kHost, kAny, kNCHW, int8);
+USE_LITE_KERNEL(affine_grid, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sequence_pool, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(rnn, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(negative, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(gru_unit, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(fetch, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(density_prior_box, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(scatter_nd_add, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(scatter_nd_add, kHost, kFloat, kNCHW, float32_int64);
+USE_LITE_KERNEL(scatter_nd_add, kHost, kFloat, kNCHW, int32_int32);
+USE_LITE_KERNEL(scatter_nd_add, kHost, kFloat, kNCHW, int32_int64);
+USE_LITE_KERNEL(scatter_nd_add, kHost, kFloat, kNCHW, int64_int32);
+USE_LITE_KERNEL(scatter_nd_add, kHost, kFloat, kNCHW, int64_int64);
+USE_LITE_KERNEL(roi_perspective_transform, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(while, kHost, kAny, kAny, def);
+USE_LITE_KERNEL(strided_slice, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(strided_slice, kHost, kFloat, kNCHW, def_int32);
+USE_LITE_KERNEL(strided_slice, kHost, kFloat, kNCHW, def_int64);
+USE_LITE_KERNEL(multiclass_nms, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(multiclass_nms2, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(multiclass_nms3, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(mul, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(mul, kARM, kInt8, kNCHW, def);
+USE_LITE_KERNEL(gather, kHost, kFloat, kNCHW, int32int32);
+USE_LITE_KERNEL(gather, kHost, kFloat, kNCHW, int64int64);
+USE_LITE_KERNEL(gather, kHost, kFloat, kNCHW, int64int32);
+USE_LITE_KERNEL(gather, kHost, kFloat, kNCHW, int32int64);
+USE_LITE_KERNEL(sequence_softmax, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(tile, kHost, kFloat, kNCHW, def_float);
+USE_LITE_KERNEL(tile, kHost, kFloat, kNCHW, def_int32);
+USE_LITE_KERNEL(tile, kHost, kFloat, kNCHW, def_int64);
+USE_LITE_KERNEL(tile, kHost, kFloat, kNCHW, def_int8);
+USE_LITE_KERNEL(tile, kHost, kFloat, kNCHW, def_bool);
+USE_LITE_KERNEL(decode_bboxes, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(slice, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(slice, kARM, kFloat, kNCHW, array_def);
+USE_LITE_KERNEL(slice, kARM, kFloat, kNCHW, float_i64_starts_ends);
+USE_LITE_KERNEL(slice, kARM, kFloat, kNCHW, array_float_i64_starts_ends);
+USE_LITE_KERNEL(slice, kARM, kFloat, kNCHW, bool_slice);
+USE_LITE_KERNEL(slice, kARM, kFloat, kNCHW, array_bool_slice);
+USE_LITE_KERNEL(slice, kARM, kFloat, kNCHW, int32_slice);
+USE_LITE_KERNEL(slice, kARM, kFloat, kNCHW, array_int32_slice);
+USE_LITE_KERNEL(slice, kARM, kFloat, kNCHW, def_int64);
+USE_LITE_KERNEL(slice, kARM, kFloat, kNCHW, array_def_int64);
+USE_LITE_KERNEL(dropout, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(fill_any_like, kHost, kAny, kNCHW, def);
+USE_LITE_KERNEL(fill_zeros_like, kHost, kAny, kNCHW, def);
+USE_LITE_KERNEL(max_pool2d_with_index, kHost, kFloat, kNCHW, fp32);
+USE_LITE_KERNEL(pad2d, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(deformable_conv, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(reduce_sum, kARM, kFloat, kNCHW, def_int32);
+USE_LITE_KERNEL(reduce_sum, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(expand_v2, kHost, kFloat, kAny, def);
+USE_LITE_KERNEL(expand_v2, kHost, kFloat, kAny, def_int32);
+USE_LITE_KERNEL(expand_v2, kHost, kFloat, kAny, def_int64);
+USE_LITE_KERNEL(affine_channel, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(crf_decoding, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(cumsum, kHost, kFloat, kAny, float32);
+USE_LITE_KERNEL(cumsum, kHost, kFloat, kAny, int32);
+USE_LITE_KERNEL(cumsum, kHost, kFloat, kAny, int64);
+USE_LITE_KERNEL(lookup_table_dequant, kARM, kAny, kNCHW, def);
+USE_LITE_KERNEL(box_coder, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(sin, kHost, kFloat, kNCHW, def);
+USE_LITE_KERNEL(fc, kARM, kFloat, kNCHW, def);
+USE_LITE_KERNEL(fc, kARM, kInt8, kNCHW, int8out);
+USE_LITE_KERNEL(fc, kARM, kInt8, kNCHW, fp32out);
+USE_LITE_KERNEL(retinanet_detection_output, kHost, kFloat, kNCHW, def);

libs/paddleocr/PaddleLite/cxx/include/paddle_use_ops.h

@@ -0,0 +1,284 @@
+#pragma once
+#include "paddle_lite_factory_helper.h"
+
+USE_LITE_OP(__xpu__logit);
+USE_LITE_OP(mean);
+USE_LITE_OP(uniform_random);
+USE_LITE_OP(equal);
+USE_LITE_OP(not_equal);
+USE_LITE_OP(less_than);
+USE_LITE_OP(less_equal);
+USE_LITE_OP(greater_than);
+USE_LITE_OP(greater_equal);
+USE_LITE_OP(fake_quantize_moving_average_abs_max);
+USE_LITE_OP(matrix_nms);
+USE_LITE_OP(lrn);
+USE_LITE_OP(axpy);
+USE_LITE_OP(inverse);
+USE_LITE_OP(scatter);
+USE_LITE_OP(__xpu__multi_encoder);
+USE_LITE_OP(gaussian_random);
+USE_LITE_OP(affine_grid);
+USE_LITE_OP(roi_align);
+USE_LITE_OP(range);
+USE_LITE_OP(sequence_topk_avg_pooling);
+USE_LITE_OP(search_group_padding);
+USE_LITE_OP(__xpu__embedding_with_eltwise_add);
+USE_LITE_OP(fake_dequantize_max_abs);
+USE_LITE_OP(strided_slice);
+USE_LITE_OP(lookup_table_dequant);
+USE_LITE_OP(sum);
+USE_LITE_OP(tile);
+USE_LITE_OP(fake_quantize_dequantize_moving_average_abs_max);
+USE_LITE_OP(search_grnn);
+USE_LITE_OP(density_prior_box);
+USE_LITE_OP(atan);
+USE_LITE_OP(lstm);
+USE_LITE_OP(fc);
+USE_LITE_OP(im2sequence);
+USE_LITE_OP(__xpu__softmax_topk);
+USE_LITE_OP(multiclass_nms);
+USE_LITE_OP(multiclass_nms2);
+USE_LITE_OP(multiclass_nms3);
+USE_LITE_OP(box_coder);
+USE_LITE_OP(sequence_reshape);
+USE_LITE_OP(sequence_conv);
+USE_LITE_OP(deformable_conv);
+USE_LITE_OP(search_aligned_mat_mul);
+USE_LITE_OP(search_fc);
+USE_LITE_OP(pad3d);
+USE_LITE_OP(dropout);
+USE_LITE_OP(calib);
+USE_LITE_OP(instance_norm);
+USE_LITE_OP(nearest_interp);
+USE_LITE_OP(bilinear_interp);
+USE_LITE_OP(stack);
+USE_LITE_OP(search_seq_depadding);
+USE_LITE_OP(cumsum);
+USE_LITE_OP(gather);
+USE_LITE_OP(sequence_unpad);
+USE_LITE_OP(pool2d);
+USE_LITE_OP(search_seq_softmax);
+USE_LITE_OP(feed);
+USE_LITE_OP(generate_proposals_v2);
+USE_LITE_OP(calib_once);
+USE_LITE_OP(pow);
+USE_LITE_OP(unstack);
+USE_LITE_OP(flip);
+USE_LITE_OP(lod_reset);
+USE_LITE_OP(beam_search_decode);
+USE_LITE_OP(io_copy_once);
+USE_LITE_OP(negative);
+USE_LITE_OP(tan);
+USE_LITE_OP(beam_search);
+USE_LITE_OP(scatter_nd_add);
+USE_LITE_OP(expand);
+USE_LITE_OP(affine_channel);
+USE_LITE_OP(sequence_mask);
+USE_LITE_OP(argsort);
+USE_LITE_OP(top_k);
+USE_LITE_OP(__xpu__fc);
+USE_LITE_OP(fill_constant_batch_size_like);
+USE_LITE_OP(unsqueeze);
+USE_LITE_OP(unsqueeze2);
+USE_LITE_OP(split);
+USE_LITE_OP(attention_padding_mask);
+USE_LITE_OP(search_attention_padding_mask);
+USE_LITE_OP(fake_quantize_range_abs_max);
+USE_LITE_OP(fake_quantize_abs_max);
+USE_LITE_OP(__xpu__multi_softmax);
+USE_LITE_OP(one_hot);
+USE_LITE_OP(max_pool2d_with_index);
+USE_LITE_OP(sin);
+USE_LITE_OP(while);
+USE_LITE_OP(ctc_align);
+USE_LITE_OP(reshape);
+USE_LITE_OP(reshape2);
+USE_LITE_OP(sequence_concat);
+USE_LITE_OP(fill_constant);
+USE_LITE_OP(flatten);
+USE_LITE_OP(flatten2);
+USE_LITE_OP(flatten_contiguous_range);
+USE_LITE_OP(is_empty);
+USE_LITE_OP(retinanet_detection_output);
+USE_LITE_OP(elementwise_sub);
+USE_LITE_OP(elementwise_add);
+USE_LITE_OP(elementwise_mul);
+USE_LITE_OP(elementwise_max);
+USE_LITE_OP(elementwise_min);
+USE_LITE_OP(elementwise_div);
+USE_LITE_OP(elementwise_floordiv);
+USE_LITE_OP(elementwise_mod);
+USE_LITE_OP(elementwise_pow);
+USE_LITE_OP(sequence_softmax);
+USE_LITE_OP(reduce_sum);
+USE_LITE_OP(reduce_prod);
+USE_LITE_OP(reduce_max);
+USE_LITE_OP(reduce_min);
+USE_LITE_OP(reduce_all);
+USE_LITE_OP(reduce_any);
+USE_LITE_OP(reduce_mean);
+USE_LITE_OP(__xpu__conv2d);
+USE_LITE_OP(lod_array_length);
+USE_LITE_OP(var_conv_2d);
+USE_LITE_OP(print);
+USE_LITE_OP(shuffle_channel);
+USE_LITE_OP(square);
+USE_LITE_OP(relu_clipped);
+USE_LITE_OP(swish);
+USE_LITE_OP(log);
+USE_LITE_OP(exp);
+USE_LITE_OP(abs);
+USE_LITE_OP(floor);
+USE_LITE_OP(hard_sigmoid);
+USE_LITE_OP(sqrt);
+USE_LITE_OP(rsqrt);
+USE_LITE_OP(softsign);
+USE_LITE_OP(gelu);
+USE_LITE_OP(hard_swish);
+USE_LITE_OP(reciprocal);
+USE_LITE_OP(mish);
+USE_LITE_OP(sigmoid);
+USE_LITE_OP(tanh);
+USE_LITE_OP(relu);
+USE_LITE_OP(leaky_relu);
+USE_LITE_OP(relu6);
+USE_LITE_OP(prelu);
+USE_LITE_OP(thresholded_relu);
+USE_LITE_OP(elu);
+USE_LITE_OP(erf);
+USE_LITE_OP(softplus);
+USE_LITE_OP(fusion_elementwise_sub_activation);
+USE_LITE_OP(fusion_elementwise_add_activation);
+USE_LITE_OP(fusion_elementwise_mul_activation);
+USE_LITE_OP(fusion_elementwise_max_activation);
+USE_LITE_OP(fusion_elementwise_min_activation);
+USE_LITE_OP(fusion_elementwise_div_activation);
+USE_LITE_OP(assign);
+USE_LITE_OP(correlation);
+USE_LITE_OP(arg_max);
+USE_LITE_OP(sign);
+USE_LITE_OP(read_from_array);
+USE_LITE_OP(logical_xor);
+USE_LITE_OP(logical_and);
+USE_LITE_OP(logical_or);
+USE_LITE_OP(logical_not);
+USE_LITE_OP(box_clip);
+USE_LITE_OP(dequantize_linear);
+USE_LITE_OP(softmax);
+USE_LITE_OP(__xpu__resnet50);
+USE_LITE_OP(transpose);
+USE_LITE_OP(transpose2);
+USE_LITE_OP(tensor_array_to_tensor);
+USE_LITE_OP(io_copy);
+USE_LITE_OP(sparse_conv2d);
+USE_LITE_OP(gather_tree);
+USE_LITE_OP(fake_channel_wise_quantize_dequantize_abs_max);
+USE_LITE_OP(increment);
+USE_LITE_OP(batch_norm);
+USE_LITE_OP(sync_batch_norm);
+USE_LITE_OP(pad2d);
+USE_LITE_OP(lookup_table_v2);
+USE_LITE_OP(unbind);
+USE_LITE_OP(distribute_fpn_proposals);
+USE_LITE_OP(fake_quantize_dequantize_abs_max);
+USE_LITE_OP(sampling_id);
+USE_LITE_OP(fpga_conv2d);
+USE_LITE_OP(index_select);
+USE_LITE_OP(match_matrix_tensor);
+USE_LITE_OP(where_index);
+USE_LITE_OP(sequence_expand_as);
+USE_LITE_OP(acos);
+USE_LITE_OP(one_hot_v2);
+USE_LITE_OP(search_seq_fc);
+USE_LITE_OP(expand_v2);
+USE_LITE_OP(topk_pooling);
+USE_LITE_OP(squeeze);
+USE_LITE_OP(squeeze2);
+USE_LITE_OP(sequence_reverse_embedding);
+USE_LITE_OP(assign_value);
+USE_LITE_OP(log_softmax);
+USE_LITE_OP(reverse);
+USE_LITE_OP(grid_sampler);
+USE_LITE_OP(bilinear_interp_v2);
+USE_LITE_OP(nearest_interp_v2);
+USE_LITE_OP(fill_zeros_like);
+USE_LITE_OP(sequence_pad);
+USE_LITE_OP(cos_sim);
+USE_LITE_OP(layout_once);
+USE_LITE_OP(subgraph);
+USE_LITE_OP(gru_unit);
+USE_LITE_OP(unfold);
+USE_LITE_OP(slice);
+USE_LITE_OP(split_lod_tensor);
+USE_LITE_OP(concat);
+USE_LITE_OP(anchor_generator);
+USE_LITE_OP(tril_triu);
+USE_LITE_OP(conditional_block);
+USE_LITE_OP(sequence_arithmetic);
+USE_LITE_OP(search_seq_arithmetic);
+USE_LITE_OP(unique_with_counts);
+USE_LITE_OP(group_norm);
+USE_LITE_OP(yolo_box);
+USE_LITE_OP(shape);
+USE_LITE_OP(matmul_v2);
+USE_LITE_OP(sequence_expand);
+USE_LITE_OP(cos);
+USE_LITE_OP(crop);
+USE_LITE_OP(layer_norm);
+USE_LITE_OP(asin);
+USE_LITE_OP(sequence_reverse);
+USE_LITE_OP(fetch);
+USE_LITE_OP(linspace);
+USE_LITE_OP(fake_channel_wise_dequantize_max_abs);
+USE_LITE_OP(rnn);
+USE_LITE_OP(__xpu__mmdnn_bid_emb_grnn_att);
+USE_LITE_OP(__xpu__mmdnn_bid_emb_grnn_att2);
+USE_LITE_OP(__xpu__mmdnn_bid_emb_att);
+USE_LITE_OP(__xpu__mmdnn_match_conv_topk);
+USE_LITE_OP(__xpu__mmdnn_merge_all);
+USE_LITE_OP(write_to_array);
+USE_LITE_OP(write_back);
+USE_LITE_OP(__xpu__sfa_head);
+USE_LITE_OP(collect_fpn_proposals);
+USE_LITE_OP(pixel_shuffle);
+USE_LITE_OP(generate_proposals);
+USE_LITE_OP(where);
+USE_LITE_OP(lookup_table);
+USE_LITE_OP(prior_box);
+USE_LITE_OP(sequence_pool_concat);
+USE_LITE_OP(conv2d);
+USE_LITE_OP(conv3d);
+USE_LITE_OP(depthwise_conv2d);
+USE_LITE_OP(polygon_box_transform);
+USE_LITE_OP(scale);
+USE_LITE_OP(roi_perspective_transform);
+USE_LITE_OP(crop_tensor);
+USE_LITE_OP(gather_nd);
+USE_LITE_OP(fill_any_like);
+USE_LITE_OP(cast);
+USE_LITE_OP(mul);
+USE_LITE_OP(quantize_linear);
+USE_LITE_OP(gru);
+USE_LITE_OP(meshgrid);
+USE_LITE_OP(__xpu__mmdnn_search_attention);
+USE_LITE_OP(__xpu__mmdnn_search_attention2);
+USE_LITE_OP(__xpu__generate_sequence);
+USE_LITE_OP(__xpu__squeeze_excitation_block);
+USE_LITE_OP(decode_bboxes);
+USE_LITE_OP(matmul);
+USE_LITE_OP(__xpu__resnet_cbam);
+USE_LITE_OP(clip);
+USE_LITE_OP(top_k_v2);
+USE_LITE_OP(__xpu__dynamic_lstm_fuse_op);
+USE_LITE_OP(norm);
+USE_LITE_OP(p_norm);
+USE_LITE_OP(crf_decoding);
+USE_LITE_OP(layout);
+USE_LITE_OP(sequence_pool);
+USE_LITE_OP(__xpu__bigru);
+USE_LITE_OP(select_input);
+USE_LITE_OP(conv2d_transpose);
+USE_LITE_OP(depthwise_conv2d_transpose);
+USE_LITE_OP(merge_lod_tensor);
+USE_LITE_OP(expand_as);

libs/paddleocr/PaddleLite/cxx/include/paddle_use_passes.h

@@ -0,0 +1,120 @@
+// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+#include "paddle_lite_factory_helper.h"  // NOLINT
+
+USE_MIR_PASS(demo);
+USE_MIR_PASS(static_kernel_pick_pass);
+USE_MIR_PASS(lite_unsqueeze2_pad3d_squeeze2_fuse_pass);
+USE_MIR_PASS(op_transformation_pass);
+USE_MIR_PASS(variable_place_inference_pass);
+USE_MIR_PASS(type_target_cast_pass);
+USE_MIR_PASS(__fpga_kernel_place_correct_pass);
+USE_MIR_PASS(opencl_kernel_place_correct_pass);
+USE_MIR_PASS(generate_program_pass);
+
+USE_MIR_PASS(io_copy_kernel_pick_pass);
+USE_MIR_PASS(argument_type_display_pass);
+USE_MIR_PASS(runtime_context_assign_pass);
+USE_MIR_PASS(graph_visualize_pass);
+
+USE_MIR_PASS(sparse_conv_detect_pass);
+USE_MIR_PASS(adaptive_1x1_pool2d_convert_global_pass);
+USE_MIR_PASS(remove_scale1_pass);
+USE_MIR_PASS(remove_tf_redundant_ops_pass);
+USE_MIR_PASS(lite_conv_bn_fuse_pass);
+USE_MIR_PASS(lite_conv_conv_fuse_pass);
+USE_MIR_PASS(lite_squeeze2_matmul_fuse_pass);
+USE_MIR_PASS(lite_reshape2_matmul_fuse_pass);
+USE_MIR_PASS(lite_matmul_fuse_pass);
+USE_MIR_PASS(lite_fc_fuse_pass);
+USE_MIR_PASS(lite_matmul_element_add_fuse_pass);
+USE_MIR_PASS(lite_shuffle_channel_fuse_pass);
+USE_MIR_PASS(lite_transpose_softmax_transpose_fuse_pass);
+USE_MIR_PASS(lite_interpolate_fuse_pass);
+USE_MIR_PASS(lite_sequence_pool_concat_fuse_pass);
+USE_MIR_PASS(identity_scale_eliminate_pass);
+USE_MIR_PASS(identity_dropout_eliminate_pass);
+USE_MIR_PASS(lite_conv_elementwise_fuse_pass);
+USE_MIR_PASS(lite_conv_activation_fuse_pass);
+USE_MIR_PASS(lite_var_conv_2d_activation_fuse_pass);
+USE_MIR_PASS(lite_match_matrix_activation_fuse_pass);
+USE_MIR_PASS(lite_scales_fuse_pass);
+USE_MIR_PASS(lite_scaleacts_fuse_pass);
+USE_MIR_PASS(lite_sequence_reverse_embedding_fuse_pass);
+USE_MIR_PASS(lite_elementwise_activation_fuse_pass);
+USE_MIR_PASS(lite_elementwise_scale_fuse_pass);
+USE_MIR_PASS(lite_conv_scale_fuse_pass);
+USE_MIR_PASS(lite_conv_elementwise_tree_fuse_pass);
+USE_MIR_PASS(lite_quant_dequant_fuse_pass);
+USE_MIR_PASS(type_precision_cast_pass);
+USE_MIR_PASS(type_layout_cast_pass);
+USE_MIR_PASS(type_layout_cast_preprocess_pass);
+USE_MIR_PASS(memory_optimize_pass);
+USE_MIR_PASS(xpu_memory_optimize_pass);
+USE_MIR_PASS(lite_inplace_fuse_pass);
+USE_MIR_PASS(multi_stream_analysis_pass);
+USE_MIR_PASS(elementwise_mul_constant_eliminate_pass);
+USE_MIR_PASS(npu_subgraph_pass);
+USE_MIR_PASS(nnadapter_subgraph_pass);
+USE_MIR_PASS(mlu_subgraph_pass);
+USE_MIR_PASS(mlu_postprocess_pass);
+USE_MIR_PASS(weight_quantization_preprocess_pass);
+USE_MIR_PASS(post_quant_dynamic_pass);
+USE_MIR_PASS(fp16_attribute_pass);
+USE_MIR_PASS(fpga_concat_fuse_pass);
+USE_MIR_PASS(quantization_parameters_propagation_pass);
+USE_MIR_PASS(quantization_parameters_removal_pass);
+USE_MIR_PASS(restrict_quantized_op_with_same_input_output_scale_pass);
+USE_MIR_PASS(control_flow_op_unused_inputs_and_outputs_eliminate_pass);
+USE_MIR_PASS(control_flow_op_shared_inputs_and_outputs_place_sync_pass);
+USE_MIR_PASS(lite_scale_activation_fuse_pass);
+USE_MIR_PASS(lite_instance_norm_activation_fuse_pass);
+USE_MIR_PASS(ssd_boxes_calc_offline_pass);
+USE_MIR_PASS(fix_mismatched_precision_pass);
+USE_MIR_PASS(lite_flatten_fc_fuse_pass);
+USE_MIR_PASS(lite_fc_prelu_fuse_pass);
+USE_MIR_PASS(lite_greater_than_cast_fuse_pass);
+USE_MIR_PASS(assign_value_calc_offline_pass);
+USE_MIR_PASS(__xpu__graph_dedup_pass);
+USE_MIR_PASS(__xpu__resnet_fuse_pass);
+USE_MIR_PASS(__xpu__resnet_cbam_fuse_pass);
+USE_MIR_PASS(__xpu__multi_encoder_fuse_pass);
+USE_MIR_PASS(__xpu__embedding_with_eltwise_add_fuse_pass);
+USE_MIR_PASS(__xpu__fc_fuse_pass);
+USE_MIR_PASS(__xpu__mmdnn_fuse_pass);
+USE_MIR_PASS(__xpu__conv2d_affine_channel_fuse_pass);
+USE_MIR_PASS(__xpu__conv2d_fuse_pass);
+USE_MIR_PASS(__xpu__sfa_head_meanstd_fuse_pass);
+USE_MIR_PASS(__xpu__sfa_head_moment_fuse_pass);
+USE_MIR_PASS(__xpu__softmax_topk_fuse_pass);
+USE_MIR_PASS(__xpu__multi_encoder_adaptive_seqlen_fuse_pass);
+USE_MIR_PASS(__xpu__multi_encoder_slice_link_fuse_pass);
+USE_MIR_PASS(__xpu__generate_sequence_fuse_pass);
+USE_MIR_PASS(__xpu__logit_fuse_pass);
+USE_MIR_PASS(__xpu__link_previous_out_max_pass);
+USE_MIR_PASS(__xpu__squeeze_excitation_fuse_pass);
+USE_MIR_PASS(__xpu__bigru_fuse_pass);
+USE_MIR_PASS(__xpu__dynamic_lstm_fuse_pass);
+USE_MIR_PASS(__xpu__multi_softmax_fuse_pass);
+USE_MIR_PASS(__xpu__max_pooling_pad_zero_detect_fuse_pass);
+USE_MIR_PASS(x86_int8_attribute_pass);
+USE_MIR_PASS(fill_range_fuse_pass);
+USE_MIR_PASS(range_calc_offline_pass);
+USE_MIR_PASS(p_norm_fill_constant_max_div_fuse_pass);
+USE_MIR_PASS(fill_constant_calc_offline_pass);
+USE_MIR_PASS(unsqueeze_calc_offline_pass);
+USE_MIR_PASS(scale_calc_offline_pass);
+USE_MIR_PASS(keepdims_convert_pass);

libs/paddleocr/build.gradle

@@ -0,0 +1,86 @@
+plugins {
+    id 'com.android.library'
+}
+import java.security.MessageDigest
+
+android {
+    namespace = "paddleocr"
+    ndkVersion '21.1.6352462'
+    compileSdk 33
+    defaultConfig {
+        minSdkVersion 24
+        targetSdkVersion 31
+
+        externalNativeBuild {
+            cmake {
+                cppFlags "-std=c++11 -frtti -fexceptions -Wno-format"
+                arguments '-DANDROID_PLATFORM=android-23', '-DANDROID_STL=c++_shared' ,"-DANDROID_ARM_NEON=TRUE"
+            }
+        }
+        ndk {
+            abiFilters "arm64-v8a", "armeabi-v7a"
+            ldLibs "jnigraphics"
+        }
+    }
+    buildTypes {
+        release {
+            minifyEnabled false
+        }
+    }
+    externalNativeBuild {
+        cmake {
+            path "src/main/cpp/CMakeLists.txt"
+            version "3.10.2"
+        }
+    }
+}
+
+dependencies {
+    implementation fileTree(include: ['*.jar'], dir: 'libs')
+    implementation project(path: ':libs:org.opencv-4.5.5')
+}
+
+/**
+ * 下载opencv4.5.5源码包 默认为4.2.0 和 AutoJS中的版本不匹配会产生冲突 这里进行修改
+ */
+def zipArchives = [
+        [
+                'src' : 'https://github.com/opencv/opencv/releases/download/4.5.5/opencv-4.5.5-android-sdk.zip',
+                'dest': 'OpenCV'
+        ]
+]
+
+/**
+ * 首次导入最好手动运行一下这个task下载依赖的包
+ */
+tasks.register('downloadAndExtractArchives', DefaultTask) {
+    doFirst {
+        println "Downloading and extracting archives including libs and models"
+    }
+    doLast {
+        // Prepare cache folder for archives
+        String cachePath = "cache"
+        if (!file("${cachePath}").exists()) {
+            mkdir "${cachePath}"
+        }
+        zipArchives.eachWithIndex { archive, index ->
+            MessageDigest messageDigest = MessageDigest.getInstance('MD5')
+            messageDigest.update(archive.src.bytes)
+            String cacheName = new BigInteger(1, messageDigest.digest()).toString(32)
+            // Download the target archive if not exists
+            boolean copyFiles = !file("${archive.dest}").exists()
+            if (!file("${cachePath}/${cacheName}.zip").exists()) {
+                ant.get(src: archive.src, dest: file("${cachePath}/${cacheName}.zip"))
+                copyFiles = true; // force to copy files from the latest archive files
+            }
+            // Extract the target archive if its dest path does not exists
+            if (copyFiles) {
+                copy {
+                    from zipTree("${cachePath}/${cacheName}.zip")
+                    into "${archive.dest}"
+                }
+            }
+        }
+    }
+}
+preBuild.dependsOn downloadAndExtractArchives

libs/paddleocr/proguard-rules.pro

@@ -0,0 +1,21 @@
+# Add project specific ProGuard rules here.
+# You can control the set of applied configuration files using the
+# proguardFiles setting in build.gradle.
+#
+# For more details, see
+#   http://developer.android.com/guide/developing/tools/proguard.html
+
+# If your project uses WebView with JS, uncomment the following
+# and specify the fully qualified class name to the JavaScript interface
+# class:
+#-keepclassmembers class fqcn.of.javascript.interface.for.webview {
+#   public *;
+#}
+
+# Uncomment this to preserve the line number information for
+# debugging stack traces.
+#-keepattributes SourceFile,LineNumberTable
+
+# If you keep the line number information, uncomment this to
+# hide the original source file name.
+#-renamesourcefileattribute SourceFile

libs/paddleocr/src/main/AndroidManifest.xml

@@ -0,0 +1,4 @@
+<?xml version="1.0" encoding="utf-8"?>
+<manifest xmlns:android="http://schemas.android.com/apk/res/android">
+
+</manifest>

libs/paddleocr/src/main/cpp/CMakeLists.txt

@@ -0,0 +1,93 @@
+# For more information about using CMake with Android Studio, read the
+# documentation: https://d.android.com/studio/projects/add-native-code.html
+
+# Sets the minimum version of CMake required to build the native library.
+
+cmake_minimum_required(VERSION 3.4.1)
+
+# Creates and names a library, sets it as either STATIC or SHARED, and provides
+# the relative paths to its source code. You can define multiple libraries, and
+# CMake builds them for you. Gradle automatically packages shared libraries with
+# your APK.
+
+set(PaddleLite_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../../../PaddleLite")
+include_directories(${PaddleLite_DIR}/cxx/include)
+
+set(OpenCV_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../../../OpenCV/OpenCV-android-sdk/sdk/native/jni")
+message(STATUS "opencv dir: ${OpenCV_DIR}")
+find_package(OpenCV REQUIRED)
+message(STATUS "OpenCV libraries: ${OpenCV_LIBS}")
+include_directories(${OpenCV_INCLUDE_DIRS})
+aux_source_directory(. SOURCES)
+set(CMAKE_CXX_FLAGS
+        "${CMAKE_CXX_FLAGS} -ffast-math -Ofast -Os"
+        )
+set(CMAKE_CXX_FLAGS
+        "${CMAKE_CXX_FLAGS} -fvisibility=hidden -fvisibility-inlines-hidden -fdata-sections -ffunction-sections"
+        )
+set(CMAKE_SHARED_LINKER_FLAGS
+        "${CMAKE_SHARED_LINKER_FLAGS} -Wl,--gc-sections -Wl,-z,nocopyreloc")
+
+add_library(
+        # Sets the name of the library.
+        Native
+        # Sets the library as a shared library.
+        SHARED
+        # Provides a relative path to your source file(s).
+        ${SOURCES})
+
+find_library(
+        # Sets the name of the path variable.
+        log-lib
+        # Specifies the name of the NDK library that you want CMake to locate.
+        log)
+
+add_library(
+        # Sets the name of the library.
+        paddle_light_api_shared
+        # Sets the library as a shared library.
+        SHARED
+        # Provides a relative path to your source file(s).
+        IMPORTED)
+
+set_target_properties(
+        # Specifies the target library.
+        paddle_light_api_shared
+        # Specifies the parameter you want to define.
+        PROPERTIES
+        IMPORTED_LOCATION
+        ${PaddleLite_DIR}/cxx/libs/${ANDROID_ABI}/libpaddle_light_api_shared.so
+        # Provides the path to the library you want to import.
+)
+
+
+# Specifies libraries CMake should link to your target library. You can link
+# multiple libraries, such as libraries you define in this build script,
+# prebuilt third-party libraries, or system libraries.
+
+target_link_libraries(
+        # Specifies the target library.
+        Native
+        paddle_light_api_shared
+        ${OpenCV_LIBS}
+        GLESv2
+        EGL
+        jnigraphics
+        ${log-lib}
+)
+
+add_custom_command(
+        TARGET Native
+        POST_BUILD
+        COMMAND
+        ${CMAKE_COMMAND} -E copy
+        ${PaddleLite_DIR}/cxx/libs/${ANDROID_ABI}/libc++_shared.so
+        ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/libc++_shared.so)
+
+add_custom_command(
+        TARGET Native
+        POST_BUILD
+        COMMAND
+        ${CMAKE_COMMAND} -E copy
+        ${PaddleLite_DIR}/cxx/libs/${ANDROID_ABI}/libpaddle_light_api_shared.so
+        ${CMAKE_LIBRARY_OUTPUT_DIRECTORY}/libpaddle_light_api_shared.so)

libs/paddleocr/src/main/cpp/common.h

@@ -0,0 +1,37 @@
+//
+// Created by fu on 4/25/18.
+//
+
+#pragma once
+#import <numeric>
+#import <vector>
+
+#ifdef __ANDROID__
+
+#include <android/log.h>
+
+#define LOG_TAG "OCR_NDK"
+
+#define LOGI(...) __android_log_print(ANDROID_LOG_INFO, LOG_TAG, __VA_ARGS__)
+#define LOGW(...) __android_log_print(ANDROID_LOG_WARN, LOG_TAG, __VA_ARGS__)
+#define LOGE(...) __android_log_print(ANDROID_LOG_ERROR, LOG_TAG, __VA_ARGS__)
+#else
+#include <stdio.h>
+#define LOGI(format, ...)                                                      \
+  fprintf(stdout, "[" LOG_TAG "]" format "\n", ##__VA_ARGS__)
+#define LOGW(format, ...)                                                      \
+  fprintf(stdout, "[" LOG_TAG "]" format "\n", ##__VA_ARGS__)
+#define LOGE(format, ...)                                                      \
+  fprintf(stderr, "[" LOG_TAG "]Error: " format "\n", ##__VA_ARGS__)
+#endif
+
+enum RETURN_CODE { RETURN_OK = 0 };
+
+enum NET_TYPE { NET_OCR = 900100, NET_OCR_INTERNAL = 991008 };
+
+template <typename T> inline T product(const std::vector<T> &vec) {
+  if (vec.empty()) {
+    return 0;
+  }
+  return std::accumulate(vec.begin(), vec.end(), 1, std::multiplies<T>());
+}

libs/paddleocr/src/main/cpp/native.cpp

@@ -0,0 +1,127 @@
+//
+// Created by fujiayi on 2020/7/5.
+//
+
+#include "native.h"
+#include "ocr_ppredictor.h"
+#include <algorithm>
+#include <paddle_api.h>
+#include <string>
+
+static paddle::lite_api::PowerMode str_to_cpu_mode(const std::string &cpu_mode);
+
+extern "C" JNIEXPORT jlong JNICALL
+Java_com_baidu_paddle_lite_ocr_OCRPredictorNative_init(
+    JNIEnv *env, jobject thiz, jstring j_det_model_path,
+    jstring j_rec_model_path, jstring j_cls_model_path, jint j_use_opencl, jint j_thread_num,
+    jstring j_cpu_mode) {
+  std::string det_model_path = jstring_to_cpp_string(env, j_det_model_path);
+  std::string rec_model_path = jstring_to_cpp_string(env, j_rec_model_path);
+  std::string cls_model_path = jstring_to_cpp_string(env, j_cls_model_path);
+  int thread_num = j_thread_num;
+  std::string cpu_mode = jstring_to_cpp_string(env, j_cpu_mode);
+  ppredictor::OCR_Config conf;
+  conf.use_opencl = j_use_opencl;
+  conf.thread_num = thread_num;
+  conf.mode = str_to_cpu_mode(cpu_mode);
+  ppredictor::OCR_PPredictor *orc_predictor =
+      new ppredictor::OCR_PPredictor{conf};
+  orc_predictor->init_from_file(det_model_path, rec_model_path, cls_model_path);
+  return reinterpret_cast<jlong>(orc_predictor);
+}
+
+/**
+ * "LITE_POWER_HIGH" convert to paddle::lite_api::LITE_POWER_HIGH
+ * @param cpu_mode
+ * @return
+ */
+static paddle::lite_api::PowerMode
+str_to_cpu_mode(const std::string &cpu_mode) {
+  static std::map<std::string, paddle::lite_api::PowerMode> cpu_mode_map{
+      {"LITE_POWER_HIGH", paddle::lite_api::LITE_POWER_HIGH},
+      {"LITE_POWER_LOW", paddle::lite_api::LITE_POWER_HIGH},
+      {"LITE_POWER_FULL", paddle::lite_api::LITE_POWER_FULL},
+      {"LITE_POWER_NO_BIND", paddle::lite_api::LITE_POWER_NO_BIND},
+      {"LITE_POWER_RAND_HIGH", paddle::lite_api::LITE_POWER_RAND_HIGH},
+      {"LITE_POWER_RAND_LOW", paddle::lite_api::LITE_POWER_RAND_LOW}};
+  std::string upper_key;
+  std::transform(cpu_mode.cbegin(), cpu_mode.cend(), upper_key.begin(),
+                 ::toupper);
+  auto index = cpu_mode_map.find(upper_key.c_str());
+  if (index == cpu_mode_map.end()) {
+    // 可能因为大小写转换后找不到 直接通过入参查找
+    index = cpu_mode_map.find(cpu_mode);
+    if (index != cpu_mode_map.end()) {
+      LOGI("find cpu_mode by &cpu_mode %s", cpu_mode.c_str());
+      return index->second;
+    }
+    LOGE("cpu_mode not found %s", upper_key.c_str());
+    return paddle::lite_api::LITE_POWER_HIGH;
+  } else {
+    return index->second;
+  }
+}
+
+extern "C" JNIEXPORT jfloatArray JNICALL
+Java_com_baidu_paddle_lite_ocr_OCRPredictorNative_forward(
+        JNIEnv *env, jobject thiz, jlong java_pointer,
+        jobject original_image,jint j_max_size_len, jint j_run_det, jint j_run_cls,
+        jint j_run_rec) {
+  LOGI("begin to run native forward");
+  if (java_pointer == 0) {
+    LOGE("JAVA pointer is NULL");
+    return cpp_array_to_jfloatarray(env, nullptr, 0);
+  }
+
+  cv::Mat origin = bitmap_to_cv_mat(env, original_image);
+  if (origin.size == nullptr) {
+    LOGE("origin bitmap cannot convert to CV Mat");
+    return cpp_array_to_jfloatarray(env, nullptr, 0);
+  }
+
+  int max_size_len = j_max_size_len;
+  int run_det = j_run_det;
+  int run_cls = j_run_cls;
+  int run_rec = j_run_rec;
+
+  ppredictor::OCR_PPredictor *ppredictor =
+      (ppredictor::OCR_PPredictor *)java_pointer;
+  std::vector<int64_t> dims_arr;
+  std::vector<ppredictor::OCRPredictResult> results =
+      ppredictor->infer_ocr(origin, max_size_len, run_det, run_cls, run_rec);
+  LOGI("infer_ocr finished with boxes %ld", results.size());
+
+  // 这里将std::vector<ppredictor::OCRPredictResult> 序列化成
+  // float数组，传输到java层再反序列化
+  std::vector<float> float_arr;
+  for (const ppredictor::OCRPredictResult &r : results) {
+    float_arr.push_back(r.points.size());
+    float_arr.push_back(r.word_index.size());
+    float_arr.push_back(r.score);
+    // add det point
+    for (const std::vector<int> &point : r.points) {
+      float_arr.push_back(point.at(0));
+      float_arr.push_back(point.at(1));
+    }
+    // add rec word idx
+    for (int index : r.word_index) {
+      float_arr.push_back(index);
+    }
+    // add cls result
+    float_arr.push_back(r.cls_label);
+    float_arr.push_back(r.cls_score);
+  }
+  return cpp_array_to_jfloatarray(env, float_arr.data(), float_arr.size());
+}
+
+extern "C" JNIEXPORT void JNICALL
+Java_com_baidu_paddle_lite_ocr_OCRPredictorNative_release(
+    JNIEnv *env, jobject thiz, jlong java_pointer) {
+  if (java_pointer == 0) {
+    LOGE("JAVA pointer is NULL");
+    return;
+  }
+  ppredictor::OCR_PPredictor *ppredictor =
+      (ppredictor::OCR_PPredictor *)java_pointer;
+  delete ppredictor;
+}

libs/paddleocr/src/main/cpp/native.h

@@ -0,0 +1,137 @@
+//
+// Created by fujiayi on 2020/7/5.
+//
+
+#pragma once
+
+#include "common.h"
+#include <android/bitmap.h>
+#include <jni.h>
+#include <opencv2/opencv.hpp>
+#include <string>
+#include <vector>
+
+inline std::string jstring_to_cpp_string(JNIEnv *env, jstring jstr) {
+  // In java, a unicode char will be encoded using 2 bytes (utf16).
+  // so jstring will contain characters utf16. std::string in c++ is
+  // essentially a string of bytes, not characters, so if we want to
+  // pass jstring from JNI to c++, we have convert utf16 to bytes.
+  if (!jstr) {
+    return "";
+  }
+  const jclass stringClass = env->GetObjectClass(jstr);
+  const jmethodID getBytes =
+      env->GetMethodID(stringClass, "getBytes", "(Ljava/lang/String;)[B");
+  const jbyteArray stringJbytes = (jbyteArray)env->CallObjectMethod(
+      jstr, getBytes, env->NewStringUTF("UTF-8"));
+
+  size_t length = (size_t)env->GetArrayLength(stringJbytes);
+  jbyte *pBytes = env->GetByteArrayElements(stringJbytes, NULL);
+
+  std::string ret = std::string(reinterpret_cast<char *>(pBytes), length);
+  env->ReleaseByteArrayElements(stringJbytes, pBytes, JNI_ABORT);
+
+  env->DeleteLocalRef(stringJbytes);
+  env->DeleteLocalRef(stringClass);
+  return ret;
+}
+
+inline jstring cpp_string_to_jstring(JNIEnv *env, std::string str) {
+  auto *data = str.c_str();
+  jclass strClass = env->FindClass("java/lang/String");
+  jmethodID strClassInitMethodID =
+      env->GetMethodID(strClass, "<init>", "([BLjava/lang/String;)V");
+
+  jbyteArray bytes = env->NewByteArray(strlen(data));
+  env->SetByteArrayRegion(bytes, 0, strlen(data),
+                          reinterpret_cast<const jbyte *>(data));
+
+  jstring encoding = env->NewStringUTF("UTF-8");
+  jstring res = (jstring)(
+      env->NewObject(strClass, strClassInitMethodID, bytes, encoding));
+
+  env->DeleteLocalRef(strClass);
+  env->DeleteLocalRef(encoding);
+  env->DeleteLocalRef(bytes);
+
+  return res;
+}
+
+inline jfloatArray cpp_array_to_jfloatarray(JNIEnv *env, const float *buf,
+                                            int64_t len) {
+  if (len == 0) {
+    return env->NewFloatArray(0);
+  }
+  jfloatArray result = env->NewFloatArray(len);
+  env->SetFloatArrayRegion(result, 0, len, buf);
+  return result;
+}
+
+inline jintArray cpp_array_to_jintarray(JNIEnv *env, const int *buf,
+                                        int64_t len) {
+  jintArray result = env->NewIntArray(len);
+  env->SetIntArrayRegion(result, 0, len, buf);
+  return result;
+}
+
+inline jbyteArray cpp_array_to_jbytearray(JNIEnv *env, const int8_t *buf,
+                                          int64_t len) {
+  jbyteArray result = env->NewByteArray(len);
+  env->SetByteArrayRegion(result, 0, len, buf);
+  return result;
+}
+
+inline jlongArray int64_vector_to_jlongarray(JNIEnv *env,
+                                             const std::vector<int64_t> &vec) {
+  jlongArray result = env->NewLongArray(vec.size());
+  jlong *buf = new jlong[vec.size()];
+  for (size_t i = 0; i < vec.size(); ++i) {
+    buf[i] = (jlong)vec[i];
+  }
+  env->SetLongArrayRegion(result, 0, vec.size(), buf);
+  delete[] buf;
+  return result;
+}
+
+inline std::vector<int64_t> jlongarray_to_int64_vector(JNIEnv *env,
+                                                       jlongArray data) {
+  int data_size = env->GetArrayLength(data);
+  jlong *data_ptr = env->GetLongArrayElements(data, nullptr);
+  std::vector<int64_t> data_vec(data_ptr, data_ptr + data_size);
+  env->ReleaseLongArrayElements(data, data_ptr, 0);
+  return data_vec;
+}
+
+inline std::vector<float> jfloatarray_to_float_vector(JNIEnv *env,
+                                                      jfloatArray data) {
+  int data_size = env->GetArrayLength(data);
+  jfloat *data_ptr = env->GetFloatArrayElements(data, nullptr);
+  std::vector<float> data_vec(data_ptr, data_ptr + data_size);
+  env->ReleaseFloatArrayElements(data, data_ptr, 0);
+  return data_vec;
+}
+
+inline cv::Mat bitmap_to_cv_mat(JNIEnv *env, jobject bitmap) {
+  AndroidBitmapInfo info;
+  int result = AndroidBitmap_getInfo(env, bitmap, &info);
+  if (result != ANDROID_BITMAP_RESULT_SUCCESS) {
+    LOGE("AndroidBitmap_getInfo failed, result: %d", result);
+    return cv::Mat{};
+  }
+  if (info.format != ANDROID_BITMAP_FORMAT_RGBA_8888) {
+    LOGE("Bitmap format is not RGBA_8888 !");
+    return cv::Mat{};
+  }
+  unsigned char *srcData = NULL;
+  AndroidBitmap_lockPixels(env, bitmap, (void **)&srcData);
+  cv::Mat mat = cv::Mat::zeros(info.height, info.width, CV_8UC4);
+  memcpy(mat.data, srcData, info.height * info.width * 4);
+  AndroidBitmap_unlockPixels(env, bitmap);
+  cv::cvtColor(mat, mat, cv::COLOR_RGBA2BGR);
+  /**
+  if (!cv::imwrite("/sdcard/1/copy.jpg", mat)){
+      LOGE("Write image failed " );
+  }
+   */
+  return mat;
+}

libs/paddleocr/src/main/cpp/ocr_clipper.hpp

@@ -0,0 +1,544 @@
+/*******************************************************************************
+*                                                                              *
+* Author    :  Angus Johnson                                                   *
+* Version   :  6.4.2                                                           *
+* Date      :  27 February 2017                                                *
+* Website   :  http://www.angusj.com                                           *
+* Copyright :  Angus Johnson 2010-2017                                         *
+*                                                                              *
+* License:                                                                     *
+* Use, modification & distribution is subject to Boost Software License Ver 1. *
+* http://www.boost.org/LICENSE_1_0.txt                                         *
+*                                                                              *
+* Attributions:                                                                *
+* The code in this library is an extension of Bala Vatti's clipping algorithm: *
+* "A generic solution to polygon clipping"                                     *
+* Communications of the ACM, Vol 35, Issue 7 (July 1992) pp 56-63.             *
+* http://portal.acm.org/citation.cfm?id=129906                                 *
+*                                                                              *
+* Computer graphics and geometric modeling: implementation and algorithms      *
+* By Max K. Agoston                                                            *
+* Springer; 1 edition (January 4, 2005)                                        *
+* http://books.google.com/books?q=vatti+clipping+agoston                       *
+*                                                                              *
+* See also:                                                                    *
+* "Polygon Offsetting by Computing Winding Numbers"                            *
+* Paper no. DETC2005-85513 pp. 565-575                                         *
+* ASME 2005 International Design Engineering Technical Conferences             *
+* and Computers and Information in Engineering Conference (IDETC/CIE2005)      *
+* September 24-28, 2005 , Long Beach, California, USA                          *
+* http://www.me.berkeley.edu/~mcmains/pubs/DAC05OffsetPolygon.pdf              *
+*                                                                              *
+*******************************************************************************/
+
+#ifndef clipper_hpp
+#define clipper_hpp
+
+#define CLIPPER_VERSION "6.4.2"
+
+// use_int32: When enabled 32bit ints are used instead of 64bit ints. This
+// improve performance but coordinate values are limited to the range +/- 46340
+//#define use_int32
+
+// use_xyz: adds a Z member to IntPoint. Adds a minor cost to perfomance.
+//#define use_xyz
+
+// use_lines: Enables line clipping. Adds a very minor cost to performance.
+#define use_lines
+
+// use_deprecated: Enables temporary support for the obsolete functions
+//#define use_deprecated
+
+#include <cstdlib>
+#include <cstring>
+#include <functional>
+#include <list>
+#include <ostream>
+#include <queue>
+#include <set>
+#include <stdexcept>
+#include <vector>
+
+namespace ClipperLib {
+
+enum ClipType { ctIntersection, ctUnion, ctDifference, ctXor };
+enum PolyType { ptSubject, ptClip };
+// By far the most widely used winding rules for polygon filling are
+// EvenOdd & NonZero (GDI, GDI+, XLib, OpenGL, Cairo, AGG, Quartz, SVG, Gr32)
+// Others rules include Positive, Negative and ABS_GTR_EQ_TWO (only in OpenGL)
+// see http://glprogramming.com/red/chapter11.html
+enum PolyFillType { pftEvenOdd, pftNonZero, pftPositive, pftNegative };
+
+#ifdef use_int32
+typedef int cInt;
+static cInt const loRange = 0x7FFF;
+static cInt const hiRange = 0x7FFF;
+#else
+typedef signed long long cInt;
+static cInt const loRange = 0x3FFFFFFF;
+static cInt const hiRange = 0x3FFFFFFFFFFFFFFFLL;
+typedef signed long long long64; // used by Int128 class
+typedef unsigned long long ulong64;
+
+#endif
+
+struct IntPoint {
+  cInt X;
+  cInt Y;
+#ifdef use_xyz
+  cInt Z;
+  IntPoint(cInt x = 0, cInt y = 0, cInt z = 0) : X(x), Y(y), Z(z){};
+#else
+
+  IntPoint(cInt x = 0, cInt y = 0) : X(x), Y(y){};
+#endif
+
+  friend inline bool operator==(const IntPoint &a, const IntPoint &b) {
+    return a.X == b.X && a.Y == b.Y;
+  }
+
+  friend inline bool operator!=(const IntPoint &a, const IntPoint &b) {
+    return a.X != b.X || a.Y != b.Y;
+  }
+};
+//------------------------------------------------------------------------------
+
+typedef std::vector<IntPoint> Path;
+typedef std::vector<Path> Paths;
+
+inline Path &operator<<(Path &poly, const IntPoint &p) {
+  poly.push_back(p);
+  return poly;
+}
+
+inline Paths &operator<<(Paths &polys, const Path &p) {
+  polys.push_back(p);
+  return polys;
+}
+
+std::ostream &operator<<(std::ostream &s, const IntPoint &p);
+
+std::ostream &operator<<(std::ostream &s, const Path &p);
+
+std::ostream &operator<<(std::ostream &s, const Paths &p);
+
+struct DoublePoint {
+  double X;
+  double Y;
+
+  DoublePoint(double x = 0, double y = 0) : X(x), Y(y) {}
+
+  DoublePoint(IntPoint ip) : X((double)ip.X), Y((double)ip.Y) {}
+};
+//------------------------------------------------------------------------------
+
+#ifdef use_xyz
+typedef void (*ZFillCallback)(IntPoint &e1bot, IntPoint &e1top, IntPoint &e2bot,
+                              IntPoint &e2top, IntPoint &pt);
+#endif
+
+enum InitOptions {
+  ioReverseSolution = 1,
+  ioStrictlySimple = 2,
+  ioPreserveCollinear = 4
+};
+enum JoinType { jtSquare, jtRound, jtMiter };
+enum EndType {
+  etClosedPolygon,
+  etClosedLine,
+  etOpenButt,
+  etOpenSquare,
+  etOpenRound
+};
+
+class PolyNode;
+
+typedef std::vector<PolyNode *> PolyNodes;
+
+class PolyNode {
+public:
+  PolyNode();
+
+  virtual ~PolyNode(){};
+  Path Contour;
+  PolyNodes Childs;
+  PolyNode *Parent;
+
+  PolyNode *GetNext() const;
+
+  bool IsHole() const;
+
+  bool IsOpen() const;
+
+  int ChildCount() const;
+
+private:
+  // PolyNode& operator =(PolyNode& other);
+  unsigned Index; // node index in Parent.Childs
+  bool m_IsOpen;
+  JoinType m_jointype;
+  EndType m_endtype;
+
+  PolyNode *GetNextSiblingUp() const;
+
+  void AddChild(PolyNode &child);
+
+  friend class Clipper; // to access Index
+  friend class ClipperOffset;
+};
+
+class PolyTree : public PolyNode {
+public:
+  ~PolyTree() { Clear(); };
+
+  PolyNode *GetFirst() const;
+
+  void Clear();
+
+  int Total() const;
+
+private:
+  // PolyTree& operator =(PolyTree& other);
+  PolyNodes AllNodes;
+
+  friend class Clipper; // to access AllNodes
+};
+
+bool Orientation(const Path &poly);
+
+double Area(const Path &poly);
+
+int PointInPolygon(const IntPoint &pt, const Path &path);
+
+void SimplifyPolygon(const Path &in_poly, Paths &out_polys,
+                     PolyFillType fillType = pftEvenOdd);
+
+void SimplifyPolygons(const Paths &in_polys, Paths &out_polys,
+                      PolyFillType fillType = pftEvenOdd);
+
+void SimplifyPolygons(Paths &polys, PolyFillType fillType = pftEvenOdd);
+
+void CleanPolygon(const Path &in_poly, Path &out_poly, double distance = 1.415);
+
+void CleanPolygon(Path &poly, double distance = 1.415);
+
+void CleanPolygons(const Paths &in_polys, Paths &out_polys,
+                   double distance = 1.415);
+
+void CleanPolygons(Paths &polys, double distance = 1.415);
+
+void MinkowskiSum(const Path &pattern, const Path &path, Paths &solution,
+                  bool pathIsClosed);
+
+void MinkowskiSum(const Path &pattern, const Paths &paths, Paths &solution,
+                  bool pathIsClosed);
+
+void MinkowskiDiff(const Path &poly1, const Path &poly2, Paths &solution);
+
+void PolyTreeToPaths(const PolyTree &polytree, Paths &paths);
+
+void ClosedPathsFromPolyTree(const PolyTree &polytree, Paths &paths);
+
+void OpenPathsFromPolyTree(PolyTree &polytree, Paths &paths);
+
+void ReversePath(Path &p);
+
+void ReversePaths(Paths &p);
+
+struct IntRect {
+  cInt left;
+  cInt top;
+  cInt right;
+  cInt bottom;
+};
+
+// enums that are used internally ...
+enum EdgeSide { esLeft = 1, esRight = 2 };
+
+// forward declarations (for stuff used internally) ...
+struct TEdge;
+struct IntersectNode;
+struct LocalMinimum;
+struct OutPt;
+struct OutRec;
+struct Join;
+
+typedef std::vector<OutRec *> PolyOutList;
+typedef std::vector<TEdge *> EdgeList;
+typedef std::vector<Join *> JoinList;
+typedef std::vector<IntersectNode *> IntersectList;
+
+//------------------------------------------------------------------------------
+
+// ClipperBase is the ancestor to the Clipper class. It should not be
+// instantiated directly. This class simply abstracts the conversion of sets of
+// polygon coordinates into edge objects that are stored in a LocalMinima list.
+class ClipperBase {
+public:
+  ClipperBase();
+
+  virtual ~ClipperBase();
+
+  virtual bool AddPath(const Path &pg, PolyType PolyTyp, bool Closed);
+
+  bool AddPaths(const Paths &ppg, PolyType PolyTyp, bool Closed);
+
+  virtual void Clear();
+
+  IntRect GetBounds();
+
+  bool PreserveCollinear() { return m_PreserveCollinear; };
+
+  void PreserveCollinear(bool value) { m_PreserveCollinear = value; };
+
+protected:
+  void DisposeLocalMinimaList();
+
+  TEdge *AddBoundsToLML(TEdge *e, bool IsClosed);
+
+  virtual void Reset();
+
+  TEdge *ProcessBound(TEdge *E, bool IsClockwise);
+
+  void InsertScanbeam(const cInt Y);
+
+  bool PopScanbeam(cInt &Y);
+
+  bool LocalMinimaPending();
+
+  bool PopLocalMinima(cInt Y, const LocalMinimum *&locMin);
+
+  OutRec *CreateOutRec();
+
+  void DisposeAllOutRecs();
+
+  void DisposeOutRec(PolyOutList::size_type index);
+
+  void SwapPositionsInAEL(TEdge *edge1, TEdge *edge2);
+
+  void DeleteFromAEL(TEdge *e);
+
+  void UpdateEdgeIntoAEL(TEdge *&e);
+
+  typedef std::vector<LocalMinimum> MinimaList;
+  MinimaList::iterator m_CurrentLM;
+  MinimaList m_MinimaList;
+
+  bool m_UseFullRange;
+  EdgeList m_edges;
+  bool m_PreserveCollinear;
+  bool m_HasOpenPaths;
+  PolyOutList m_PolyOuts;
+  TEdge *m_ActiveEdges;
+
+  typedef std::priority_queue<cInt> ScanbeamList;
+  ScanbeamList m_Scanbeam;
+};
+//------------------------------------------------------------------------------
+
+class Clipper : public virtual ClipperBase {
+public:
+  Clipper(int initOptions = 0);
+
+  bool Execute(ClipType clipType, Paths &solution,
+               PolyFillType fillType = pftEvenOdd);
+
+  bool Execute(ClipType clipType, Paths &solution, PolyFillType subjFillType,
+               PolyFillType clipFillType);
+
+  bool Execute(ClipType clipType, PolyTree &polytree,
+               PolyFillType fillType = pftEvenOdd);
+
+  bool Execute(ClipType clipType, PolyTree &polytree, PolyFillType subjFillType,
+               PolyFillType clipFillType);
+
+  bool ReverseSolution() { return m_ReverseOutput; };
+
+  void ReverseSolution(bool value) { m_ReverseOutput = value; };
+
+  bool StrictlySimple() { return m_StrictSimple; };
+
+  void StrictlySimple(bool value) { m_StrictSimple = value; };
+// set the callback function for z value filling on intersections (otherwise Z
+// is 0)
+#ifdef use_xyz
+  void ZFillFunction(ZFillCallback zFillFunc);
+#endif
+protected:
+  virtual bool ExecuteInternal();
+
+private:
+  JoinList m_Joins;
+  JoinList m_GhostJoins;
+  IntersectList m_IntersectList;
+  ClipType m_ClipType;
+  typedef std::list<cInt> MaximaList;
+  MaximaList m_Maxima;
+  TEdge *m_SortedEdges;
+  bool m_ExecuteLocked;
+  PolyFillType m_ClipFillType;
+  PolyFillType m_SubjFillType;
+  bool m_ReverseOutput;
+  bool m_UsingPolyTree;
+  bool m_StrictSimple;
+#ifdef use_xyz
+  ZFillCallback m_ZFill; // custom callback
+#endif
+
+  void SetWindingCount(TEdge &edge);
+
+  bool IsEvenOddFillType(const TEdge &edge) const;
+
+  bool IsEvenOddAltFillType(const TEdge &edge) const;
+
+  void InsertLocalMinimaIntoAEL(const cInt botY);
+
+  void InsertEdgeIntoAEL(TEdge *edge, TEdge *startEdge);
+
+  void AddEdgeToSEL(TEdge *edge);
+
+  bool PopEdgeFromSEL(TEdge *&edge);
+
+  void CopyAELToSEL();
+
+  void DeleteFromSEL(TEdge *e);
+
+  void SwapPositionsInSEL(TEdge *edge1, TEdge *edge2);
+
+  bool IsContributing(const TEdge &edge) const;
+
+  bool IsTopHorz(const cInt XPos);
+
+  void DoMaxima(TEdge *e);
+
+  void ProcessHorizontals();
+
+  void ProcessHorizontal(TEdge *horzEdge);
+
+  void AddLocalMaxPoly(TEdge *e1, TEdge *e2, const IntPoint &pt);
+
+  OutPt *AddLocalMinPoly(TEdge *e1, TEdge *e2, const IntPoint &pt);
+
+  OutRec *GetOutRec(int idx);
+
+  void AppendPolygon(TEdge *e1, TEdge *e2);
+
+  void IntersectEdges(TEdge *e1, TEdge *e2, IntPoint &pt);
+
+  OutPt *AddOutPt(TEdge *e, const IntPoint &pt);
+
+  OutPt *GetLastOutPt(TEdge *e);
+
+  bool ProcessIntersections(const cInt topY);
+
+  void BuildIntersectList(const cInt topY);
+
+  void ProcessIntersectList();
+
+  void ProcessEdgesAtTopOfScanbeam(const cInt topY);
+
+  void BuildResult(Paths &polys);
+
+  void BuildResult2(PolyTree &polytree);
+
+  void SetHoleState(TEdge *e, OutRec *outrec);
+
+  void DisposeIntersectNodes();
+
+  bool FixupIntersectionOrder();
+
+  void FixupOutPolygon(OutRec &outrec);
+
+  void FixupOutPolyline(OutRec &outrec);
+
+  bool IsHole(TEdge *e);
+
+  bool FindOwnerFromSplitRecs(OutRec &outRec, OutRec *&currOrfl);
+
+  void FixHoleLinkage(OutRec &outrec);
+
+  void AddJoin(OutPt *op1, OutPt *op2, const IntPoint offPt);
+
+  void ClearJoins();
+
+  void ClearGhostJoins();
+
+  void AddGhostJoin(OutPt *op, const IntPoint offPt);
+
+  bool JoinPoints(Join *j, OutRec *outRec1, OutRec *outRec2);
+
+  void JoinCommonEdges();
+
+  void DoSimplePolygons();
+
+  void FixupFirstLefts1(OutRec *OldOutRec, OutRec *NewOutRec);
+
+  void FixupFirstLefts2(OutRec *InnerOutRec, OutRec *OuterOutRec);
+
+  void FixupFirstLefts3(OutRec *OldOutRec, OutRec *NewOutRec);
+
+#ifdef use_xyz
+  void SetZ(IntPoint &pt, TEdge &e1, TEdge &e2);
+#endif
+};
+//------------------------------------------------------------------------------
+
+class ClipperOffset {
+public:
+  ClipperOffset(double miterLimit = 2.0, double roundPrecision = 0.25);
+
+  ~ClipperOffset();
+
+  void AddPath(const Path &path, JoinType joinType, EndType endType);
+
+  void AddPaths(const Paths &paths, JoinType joinType, EndType endType);
+
+  void Execute(Paths &solution, double delta);
+
+  void Execute(PolyTree &solution, double delta);
+
+  void Clear();
+
+  double MiterLimit;
+  double ArcTolerance;
+
+private:
+  Paths m_destPolys;
+  Path m_srcPoly;
+  Path m_destPoly;
+  std::vector<DoublePoint> m_normals;
+  double m_delta, m_sinA, m_sin, m_cos;
+  double m_miterLim, m_StepsPerRad;
+  IntPoint m_lowest;
+  PolyNode m_polyNodes;
+
+  void FixOrientations();
+
+  void DoOffset(double delta);
+
+  void OffsetPoint(int j, int &k, JoinType jointype);
+
+  void DoSquare(int j, int k);
+
+  void DoMiter(int j, int k, double r);
+
+  void DoRound(int j, int k);
+};
+//------------------------------------------------------------------------------
+
+class clipperException : public std::exception {
+public:
+  clipperException(const char *description) : m_descr(description) {}
+
+  virtual ~clipperException() throw() {}
+
+  virtual const char *what() const throw() { return m_descr.c_str(); }
+
+private:
+  std::string m_descr;
+};
+//------------------------------------------------------------------------------
+
+} // ClipperLib namespace
+
+#endif // clipper_hpp

libs/paddleocr/src/main/cpp/ocr_cls_process.cpp

@@ -0,0 +1,46 @@
+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "ocr_cls_process.h"
+#include <cmath>
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <iostream>
+#include <vector>
+
+const std::vector<int> CLS_IMAGE_SHAPE = {3, 48, 192};
+
+cv::Mat cls_resize_img(const cv::Mat &img) {
+  int imgC = CLS_IMAGE_SHAPE[0];
+  int imgW = CLS_IMAGE_SHAPE[2];
+  int imgH = CLS_IMAGE_SHAPE[1];
+
+  float ratio = float(img.cols) / float(img.rows);
+  int resize_w = 0;
+  if (ceilf(imgH * ratio) > imgW)
+    resize_w = imgW;
+  else
+    resize_w = int(ceilf(imgH * ratio));
+
+  cv::Mat resize_img;
+  cv::resize(img, resize_img, cv::Size(resize_w, imgH), 0.f, 0.f,
+             cv::INTER_CUBIC);
+
+  if (resize_w < imgW) {
+    cv::copyMakeBorder(resize_img, resize_img, 0, 0, 0, int(imgW - resize_w),
+                       cv::BORDER_CONSTANT, {0, 0, 0});
+  }
+  return resize_img;
+}

libs/paddleocr/src/main/cpp/ocr_cls_process.h

@@ -0,0 +1,23 @@
+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#pragma once
+
+#include "common.h"
+#include <opencv2/opencv.hpp>
+#include <vector>
+
+extern const std::vector<int> CLS_IMAGE_SHAPE;
+
+cv::Mat cls_resize_img(const cv::Mat &img);

libs/paddleocr/src/main/cpp/ocr_crnn_process.cpp

@@ -0,0 +1,144 @@
+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "ocr_crnn_process.h"
+#include <cmath>
+#include <cstring>
+#include <fstream>
+#include <iostream>
+#include <iostream>
+#include <vector>
+
+const std::string CHARACTER_TYPE = "ch";
+const int MAX_DICT_LENGTH = 6624;
+const std::vector<int> REC_IMAGE_SHAPE = {3, 48, 320};
+
+static cv::Mat crnn_resize_norm_img(cv::Mat img, float wh_ratio) {
+  int imgC = REC_IMAGE_SHAPE[0];
+  int imgH = REC_IMAGE_SHAPE[1];
+  int imgW = REC_IMAGE_SHAPE[2];
+
+  if (CHARACTER_TYPE == "ch")
+    imgW = int(32 * wh_ratio);
+
+  float ratio = float(img.cols) / float(img.rows);
+  int resize_w = 0;
+  if (ceilf(imgH * ratio) > imgW)
+    resize_w = imgW;
+  else
+    resize_w = int(ceilf(imgH * ratio));
+  cv::Mat resize_img;
+  cv::resize(img, resize_img, cv::Size(resize_w, imgH), 0.f, 0.f,
+             cv::INTER_CUBIC);
+
+  resize_img.convertTo(resize_img, CV_32FC3, 1 / 255.f);
+
+  for (int h = 0; h < resize_img.rows; h++) {
+    for (int w = 0; w < resize_img.cols; w++) {
+      resize_img.at<cv::Vec3f>(h, w)[0] =
+          (resize_img.at<cv::Vec3f>(h, w)[0] - 0.5) * 2;
+      resize_img.at<cv::Vec3f>(h, w)[1] =
+          (resize_img.at<cv::Vec3f>(h, w)[1] - 0.5) * 2;
+      resize_img.at<cv::Vec3f>(h, w)[2] =
+          (resize_img.at<cv::Vec3f>(h, w)[2] - 0.5) * 2;
+    }
+  }
+
+  cv::Mat dist;
+  cv::copyMakeBorder(resize_img, dist, 0, 0, 0, int(imgW - resize_w),
+                     cv::BORDER_CONSTANT, {0, 0, 0});
+
+  return dist;
+}
+
+cv::Mat crnn_resize_img(const cv::Mat &img, float wh_ratio) {
+  int imgC = REC_IMAGE_SHAPE[0];
+  int imgH = REC_IMAGE_SHAPE[1];
+  int imgW = REC_IMAGE_SHAPE[2];
+
+  imgW = int(imgH * wh_ratio);
+
+  float ratio = float(img.cols) / float(img.rows);
+  int resize_w = 0;
+  if (ceilf(imgH * ratio) > imgW)
+    resize_w = imgW;
+  else
+    resize_w = int(ceilf(imgH * ratio));
+  cv::Mat resize_img;
+  cv::resize(img, resize_img, cv::Size(resize_w, imgH), 0.f, 0.f,
+             cv::INTER_LINEAR);
+  cv::copyMakeBorder(resize_img, resize_img, 0, 0, 0,
+                     int(imgW - resize_img.cols), cv::BORDER_CONSTANT,
+                     {127, 127, 127});
+  return resize_img;
+}
+
+cv::Mat get_rotate_crop_image(const cv::Mat &srcimage,
+                              const std::vector<std::vector<int>> &box) {
+
+  std::vector<std::vector<int>> points = box;
+
+  int x_collect[4] = {box[0][0], box[1][0], box[2][0], box[3][0]};
+  int y_collect[4] = {box[0][1], box[1][1], box[2][1], box[3][1]};
+  int left = int(*std::min_element(x_collect, x_collect + 4));
+  int right = int(*std::max_element(x_collect, x_collect + 4));
+  int top = int(*std::min_element(y_collect, y_collect + 4));
+  int bottom = int(*std::max_element(y_collect, y_collect + 4));
+
+  cv::Mat img_crop;
+  srcimage(cv::Rect(left, top, right - left, bottom - top)).copyTo(img_crop);
+
+  for (int i = 0; i < points.size(); i++) {
+    points[i][0] -= left;
+    points[i][1] -= top;
+  }
+
+  int img_crop_width = int(sqrt(pow(points[0][0] - points[1][0], 2) +
+                                pow(points[0][1] - points[1][1], 2)));
+  int img_crop_height = int(sqrt(pow(points[0][0] - points[3][0], 2) +
+                                 pow(points[0][1] - points[3][1], 2)));
+
+  cv::Point2f pts_std[4];
+  pts_std[0] = cv::Point2f(0., 0.);
+  pts_std[1] = cv::Point2f(img_crop_width, 0.);
+  pts_std[2] = cv::Point2f(img_crop_width, img_crop_height);
+  pts_std[3] = cv::Point2f(0.f, img_crop_height);
+
+  cv::Point2f pointsf[4];
+  pointsf[0] = cv::Point2f(points[0][0], points[0][1]);
+  pointsf[1] = cv::Point2f(points[1][0], points[1][1]);
+  pointsf[2] = cv::Point2f(points[2][0], points[2][1]);
+  pointsf[3] = cv::Point2f(points[3][0], points[3][1]);
+
+  cv::Mat M = cv::getPerspectiveTransform(pointsf, pts_std);
+
+  cv::Mat dst_img;
+  cv::warpPerspective(img_crop, dst_img, M,
+                      cv::Size(img_crop_width, img_crop_height),
+                      cv::BORDER_REPLICATE);
+
+  if (float(dst_img.rows) >= float(dst_img.cols) * 1.5) {
+    /*
+    cv::Mat srcCopy = cv::Mat(dst_img.rows, dst_img.cols, dst_img.depth());
+    cv::transpose(dst_img, srcCopy);
+    cv::flip(srcCopy, srcCopy, 0);
+    return srcCopy;
+    */
+    cv::transpose(dst_img, dst_img);
+    cv::flip(dst_img, dst_img, 0);
+    return dst_img;
+  } else {
+    return dst_img;
+  }
+}

libs/paddleocr/src/main/cpp/ocr_crnn_process.h

@@ -0,0 +1,20 @@
+//
+// Created by fujiayi on 2020/7/3.
+//
+#pragma once
+
+#include "common.h"
+#include <opencv2/opencv.hpp>
+#include <vector>
+
+extern const std::vector<int> REC_IMAGE_SHAPE;
+
+cv::Mat get_rotate_crop_image(const cv::Mat &srcimage,
+                              const std::vector<std::vector<int>> &box);
+
+cv::Mat crnn_resize_img(const cv::Mat &img, float wh_ratio);
+
+template <class ForwardIterator>
+inline size_t argmax(ForwardIterator first, ForwardIterator last) {
+  return std::distance(first, std::max_element(first, last));
+}

libs/paddleocr/src/main/cpp/ocr_db_post_process.cpp

@@ -0,0 +1,342 @@
+// Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "ocr_clipper.hpp"
+#include "opencv2/core.hpp"
+#include "opencv2/imgcodecs.hpp"
+#include "opencv2/imgproc.hpp"
+#include <iostream>
+#include <math.h>
+#include <vector>
+
+static void getcontourarea(float **box, float unclip_ratio, float &distance) {
+  int pts_num = 4;
+  float area = 0.0f;
+  float dist = 0.0f;
+  for (int i = 0; i < pts_num; i++) {
+    area += box[i][0] * box[(i + 1) % pts_num][1] -
+            box[i][1] * box[(i + 1) % pts_num][0];
+    dist += sqrtf((box[i][0] - box[(i + 1) % pts_num][0]) *
+                      (box[i][0] - box[(i + 1) % pts_num][0]) +
+                  (box[i][1] - box[(i + 1) % pts_num][1]) *
+                      (box[i][1] - box[(i + 1) % pts_num][1]));
+  }
+  area = fabs(float(area / 2.0));
+
+  distance = area * unclip_ratio / dist;
+}
+
+static cv::RotatedRect unclip(float **box) {
+  float unclip_ratio = 2.0;
+  float distance = 1.0;
+
+  getcontourarea(box, unclip_ratio, distance);
+
+  ClipperLib::ClipperOffset offset;
+  ClipperLib::Path p;
+  p << ClipperLib::IntPoint(int(box[0][0]), int(box[0][1]))
+    << ClipperLib::IntPoint(int(box[1][0]), int(box[1][1]))
+    << ClipperLib::IntPoint(int(box[2][0]), int(box[2][1]))
+    << ClipperLib::IntPoint(int(box[3][0]), int(box[3][1]));
+  offset.AddPath(p, ClipperLib::jtRound, ClipperLib::etClosedPolygon);
+
+  ClipperLib::Paths soln;
+  offset.Execute(soln, distance);
+  std::vector<cv::Point2f> points;
+
+  for (int j = 0; j < soln.size(); j++) {
+    for (int i = 0; i < soln[soln.size() - 1].size(); i++) {
+      points.emplace_back(soln[j][i].X, soln[j][i].Y);
+    }
+  }
+  cv::RotatedRect res = cv::minAreaRect(points);
+
+  return res;
+}
+
+static float **Mat2Vec(cv::Mat mat) {
+  auto **array = new float *[mat.rows];
+  for (int i = 0; i < mat.rows; ++i) {
+    array[i] = new float[mat.cols];
+  }
+  for (int i = 0; i < mat.rows; ++i) {
+    for (int j = 0; j < mat.cols; ++j) {
+      array[i][j] = mat.at<float>(i, j);
+    }
+  }
+
+  return array;
+}
+
+static void quickSort(float **s, int l, int r) {
+  if (l < r) {
+    int i = l, j = r;
+    float x = s[l][0];
+    float *xp = s[l];
+    while (i < j) {
+      while (i < j && s[j][0] >= x) {
+        j--;
+      }
+      if (i < j) {
+        std::swap(s[i++], s[j]);
+      }
+      while (i < j && s[i][0] < x) {
+        i++;
+      }
+      if (i < j) {
+        std::swap(s[j--], s[i]);
+      }
+    }
+    s[i] = xp;
+    quickSort(s, l, i - 1);
+    quickSort(s, i + 1, r);
+  }
+}
+
+static void quickSort_vector(std::vector<std::vector<int>> &box, int l, int r,
+                             int axis) {
+  if (l < r) {
+    int i = l, j = r;
+    int x = box[l][axis];
+    std::vector<int> xp(box[l]);
+    while (i < j) {
+      while (i < j && box[j][axis] >= x) {
+        j--;
+      }
+      if (i < j) {
+        std::swap(box[i++], box[j]);
+      }
+      while (i < j && box[i][axis] < x) {
+        i++;
+      }
+      if (i < j) {
+        std::swap(box[j--], box[i]);
+      }
+    }
+    box[i] = xp;
+    quickSort_vector(box, l, i - 1, axis);
+    quickSort_vector(box, i + 1, r, axis);
+  }
+}
+
+static std::vector<std::vector<int>>
+order_points_clockwise(std::vector<std::vector<int>> pts) {
+  std::vector<std::vector<int>> box = pts;
+  quickSort_vector(box, 0, int(box.size() - 1), 0);
+  std::vector<std::vector<int>> leftmost = {box[0], box[1]};
+  std::vector<std::vector<int>> rightmost = {box[2], box[3]};
+
+  if (leftmost[0][1] > leftmost[1][1]) {
+    std::swap(leftmost[0], leftmost[1]);
+  }
+
+  if (rightmost[0][1] > rightmost[1][1]) {
+    std::swap(rightmost[0], rightmost[1]);
+  }
+
+  std::vector<std::vector<int>> rect = {leftmost[0], rightmost[0], rightmost[1],
+                                        leftmost[1]};
+  return rect;
+}
+
+static float **get_mini_boxes(cv::RotatedRect box, float &ssid) {
+  ssid = box.size.width >= box.size.height ? box.size.height : box.size.width;
+
+  cv::Mat points;
+  cv::boxPoints(box, points);
+  // sorted box points
+  auto array = Mat2Vec(points);
+  quickSort(array, 0, 3);
+
+  float *idx1 = array[0], *idx2 = array[1], *idx3 = array[2], *idx4 = array[3];
+  if (array[3][1] <= array[2][1]) {
+    idx2 = array[3];
+    idx3 = array[2];
+  } else {
+    idx2 = array[2];
+    idx3 = array[3];
+  }
+  if (array[1][1] <= array[0][1]) {
+    idx1 = array[1];
+    idx4 = array[0];
+  } else {
+    idx1 = array[0];
+    idx4 = array[1];
+  }
+
+  array[0] = idx1;
+  array[1] = idx2;
+  array[2] = idx3;
+  array[3] = idx4;
+
+  return array;
+}
+
+template <class T> T clamp(T x, T min, T max) {
+  if (x > max) {
+    return max;
+  }
+  if (x < min) {
+    return min;
+  }
+  return x;
+}
+
+static float clampf(float x, float min, float max) {
+  if (x > max)
+    return max;
+  if (x < min)
+    return min;
+  return x;
+}
+
+float box_score_fast(float **box_array, cv::Mat pred) {
+  auto array = box_array;
+  int width = pred.cols;
+  int height = pred.rows;
+
+  float box_x[4] = {array[0][0], array[1][0], array[2][0], array[3][0]};
+  float box_y[4] = {array[0][1], array[1][1], array[2][1], array[3][1]};
+
+  int xmin = clamp(int(std::floorf(*(std::min_element(box_x, box_x + 4)))), 0,
+                   width - 1);
+  int xmax = clamp(int(std::ceilf(*(std::max_element(box_x, box_x + 4)))), 0,
+                   width - 1);
+  int ymin = clamp(int(std::floorf(*(std::min_element(box_y, box_y + 4)))), 0,
+                   height - 1);
+  int ymax = clamp(int(std::ceilf(*(std::max_element(box_y, box_y + 4)))), 0,
+                   height - 1);
+
+  cv::Mat mask;
+  mask = cv::Mat::zeros(ymax - ymin + 1, xmax - xmin + 1, CV_8UC1);
+
+  cv::Point root_point[4];
+  root_point[0] = cv::Point(int(array[0][0]) - xmin, int(array[0][1]) - ymin);
+  root_point[1] = cv::Point(int(array[1][0]) - xmin, int(array[1][1]) - ymin);
+  root_point[2] = cv::Point(int(array[2][0]) - xmin, int(array[2][1]) - ymin);
+  root_point[3] = cv::Point(int(array[3][0]) - xmin, int(array[3][1]) - ymin);
+  const cv::Point *ppt[1] = {root_point};
+  int npt[] = {4};
+  cv::fillPoly(mask, ppt, npt, 1, cv::Scalar(1));
+
+  cv::Mat croppedImg;
+  pred(cv::Rect(xmin, ymin, xmax - xmin + 1, ymax - ymin + 1))
+      .copyTo(croppedImg);
+
+  auto score = cv::mean(croppedImg, mask)[0];
+  return score;
+}
+
+std::vector<std::vector<std::vector<int>>>
+boxes_from_bitmap(const cv::Mat &pred, const cv::Mat &bitmap) {
+  const int min_size = 3;
+  const int max_candidates = 1000;
+  const float box_thresh = 0.5;
+
+  int width = bitmap.cols;
+  int height = bitmap.rows;
+
+  std::vector<std::vector<cv::Point>> contours;
+  std::vector<cv::Vec4i> hierarchy;
+
+  cv::findContours(bitmap, contours, hierarchy, cv::RETR_LIST,
+                   cv::CHAIN_APPROX_SIMPLE);
+
+  int num_contours =
+      contours.size() >= max_candidates ? max_candidates : contours.size();
+
+  std::vector<std::vector<std::vector<int>>> boxes;
+
+  for (int _i = 0; _i < num_contours; _i++) {
+    float ssid;
+    cv::RotatedRect box = cv::minAreaRect(contours[_i]);
+    auto array = get_mini_boxes(box, ssid);
+
+    auto box_for_unclip = array;
+    // end get_mini_box
+
+    if (ssid < min_size) {
+      continue;
+    }
+
+    float score;
+    score = box_score_fast(array, pred);
+    // end box_score_fast
+    if (score < box_thresh) {
+      continue;
+    }
+
+    // start for unclip
+    cv::RotatedRect points = unclip(box_for_unclip);
+    // end for unclip
+
+    cv::RotatedRect clipbox = points;
+    auto cliparray = get_mini_boxes(clipbox, ssid);
+
+    if (ssid < min_size + 2)
+      continue;
+
+    int dest_width = pred.cols;
+    int dest_height = pred.rows;
+    std::vector<std::vector<int>> intcliparray;
+
+    for (int num_pt = 0; num_pt < 4; num_pt++) {
+      std::vector<int> a{int(clampf(roundf(cliparray[num_pt][0] / float(width) *
+                                           float(dest_width)),
+                                    0, float(dest_width))),
+                         int(clampf(roundf(cliparray[num_pt][1] /
+                                           float(height) * float(dest_height)),
+                                    0, float(dest_height)))};
+      intcliparray.emplace_back(std::move(a));
+    }
+    boxes.emplace_back(std::move(intcliparray));
+
+  } // end for
+  return boxes;
+}
+
+int _max(int a, int b) { return a >= b ? a : b; }
+
+int _min(int a, int b) { return a >= b ? b : a; }
+
+std::vector<std::vector<std::vector<int>>>
+filter_tag_det_res(const std::vector<std::vector<std::vector<int>>> &o_boxes,
+                   float ratio_h, float ratio_w, const cv::Mat &srcimg) {
+  int oriimg_h = srcimg.rows;
+  int oriimg_w = srcimg.cols;
+  std::vector<std::vector<std::vector<int>>> boxes{o_boxes};
+  std::vector<std::vector<std::vector<int>>> root_points;
+  for (int n = 0; n < boxes.size(); n++) {
+    boxes[n] = order_points_clockwise(boxes[n]);
+    for (int m = 0; m < boxes[0].size(); m++) {
+      boxes[n][m][0] /= ratio_w;
+      boxes[n][m][1] /= ratio_h;
+
+      boxes[n][m][0] = int(_min(_max(boxes[n][m][0], 0), oriimg_w - 1));
+      boxes[n][m][1] = int(_min(_max(boxes[n][m][1], 0), oriimg_h - 1));
+    }
+  }
+
+  for (int n = 0; n < boxes.size(); n++) {
+    int rect_width, rect_height;
+    rect_width = int(sqrt(pow(boxes[n][0][0] - boxes[n][1][0], 2) +
+                          pow(boxes[n][0][1] - boxes[n][1][1], 2)));
+    rect_height = int(sqrt(pow(boxes[n][0][0] - boxes[n][3][0], 2) +
+                           pow(boxes[n][0][1] - boxes[n][3][1], 2)));
+    if (rect_width <= 10 || rect_height <= 10)
+      continue;
+    root_points.push_back(boxes[n]);
+  }
+  return root_points;
+}

libs/paddleocr/src/main/cpp/ocr_db_post_process.h

@@ -0,0 +1,13 @@
+//
+// Created by fujiayi on 2020/7/2.
+//
+#pragma once
+#include <opencv2/opencv.hpp>
+#include <vector>
+
+std::vector<std::vector<std::vector<int>>>
+boxes_from_bitmap(const cv::Mat &pred, const cv::Mat &bitmap);
+
+std::vector<std::vector<std::vector<int>>>
+filter_tag_det_res(const std::vector<std::vector<std::vector<int>>> &o_boxes,
+                   float ratio_h, float ratio_w, const cv::Mat &srcimg);

libs/paddleocr/src/main/cpp/ocr_ppredictor.cpp

@@ -0,0 +1,350 @@
+//
+// Created by fujiayi on 2020/7/1.
+//
+
+#include "ocr_ppredictor.h"
+#include "common.h"
+#include "ocr_cls_process.h"
+#include "ocr_crnn_process.h"
+#include "ocr_db_post_process.h"
+#include "preprocess.h"
+
+namespace ppredictor {
+
+OCR_PPredictor::OCR_PPredictor(const OCR_Config &config) : _config(config) {}
+
+int OCR_PPredictor::init(const std::string &det_model_content,
+                         const std::string &rec_model_content,
+                         const std::string &cls_model_content) {
+  _det_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.use_opencl,_config.thread_num, NET_OCR, _config.mode});
+  _det_predictor->init_nb(det_model_content);
+
+  _rec_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.use_opencl,_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _rec_predictor->init_nb(rec_model_content);
+
+  _cls_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.use_opencl,_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _cls_predictor->init_nb(cls_model_content);
+  return RETURN_OK;
+}
+
+int OCR_PPredictor::init_from_file(const std::string &det_model_path,
+                                   const std::string &rec_model_path,
+                                   const std::string &cls_model_path) {
+  _det_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.use_opencl, _config.thread_num, NET_OCR, _config.mode});
+  _det_predictor->init_from_file(det_model_path);
+
+  _rec_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.use_opencl,_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _rec_predictor->init_from_file(rec_model_path);
+
+  _cls_predictor = std::unique_ptr<PPredictor>(
+      new PPredictor{_config.use_opencl,_config.thread_num, NET_OCR_INTERNAL, _config.mode});
+  _cls_predictor->init_from_file(cls_model_path);
+  return RETURN_OK;
+}
+/**
+ * for debug use, show result of First Step
+ * @param filter_boxes
+ * @param boxes
+ * @param srcimg
+ */
+static void
+visual_img(const std::vector<std::vector<std::vector<int>>> &filter_boxes,
+           const std::vector<std::vector<std::vector<int>>> &boxes,
+           const cv::Mat &srcimg) {
+  // visualization
+  cv::Point rook_points[filter_boxes.size()][4];
+  for (int n = 0; n < filter_boxes.size(); n++) {
+    for (int m = 0; m < filter_boxes[0].size(); m++) {
+      rook_points[n][m] =
+          cv::Point(int(filter_boxes[n][m][0]), int(filter_boxes[n][m][1]));
+    }
+  }
+
+  cv::Mat img_vis;
+  srcimg.copyTo(img_vis);
+  for (int n = 0; n < boxes.size(); n++) {
+    const cv::Point *ppt[1] = {rook_points[n]};
+    int npt[] = {4};
+    cv::polylines(img_vis, ppt, npt, 1, 1, CV_RGB(0, 255, 0), 2, 8, 0);
+  }
+  // 调试用，自行替换需要修改的路径
+  cv::imwrite("/sdcard/1/vis.png", img_vis);
+}
+
+std::vector<OCRPredictResult>
+OCR_PPredictor::infer_ocr(cv::Mat &origin,int max_size_len, int run_det, int run_cls, int run_rec) {
+  LOGI("ocr cpp start *****************");
+  LOGI("ocr cpp det: %d, cls: %d, rec: %d", run_det, run_cls, run_rec);
+  std::vector<OCRPredictResult> ocr_results;
+  if (run_det || run_rec) {
+    infer_det(origin, max_size_len, ocr_results);
+  }
+  if (run_rec) {
+    if (ocr_results.empty()) {
+      OCRPredictResult res;
+      ocr_results.emplace_back(std::move(res));
+    }
+    for (auto & ocr_result : ocr_results) {
+      infer_rec(origin, run_cls, ocr_result);
+    }
+  }
+  if (run_cls) {
+    ClsPredictResult cls_res = infer_cls(origin);
+    OCRPredictResult res;
+    res.cls_score = cls_res.cls_score;
+    res.cls_label = cls_res.cls_label;
+    ocr_results.push_back(res);
+  }
+
+  LOGI("ocr cpp end *****************");
+  return ocr_results;
+}
+
+cv::Mat DetResizeImg(const cv::Mat img, int max_size_len,
+                     std::vector<float> &ratio_hw) {
+  int w = img.cols;
+  int h = img.rows;
+
+  float ratio = 1.f;
+  int max_wh = w >= h ? w : h;
+  if (max_wh > max_size_len) {
+    if (h > w) {
+      ratio = static_cast<float>(max_size_len) / static_cast<float>(h);
+    } else {
+      ratio = static_cast<float>(max_size_len) / static_cast<float>(w);
+    }
+  }
+
+  int resize_h = static_cast<int>(float(h) * ratio);
+  int resize_w = static_cast<int>(float(w) * ratio);
+  int shapeSize = 32;
+  if (resize_h % shapeSize == 0)
+    resize_h = resize_h;
+  else if (resize_h / shapeSize < 1 + 1e-5)
+    resize_h = shapeSize;
+  else
+    resize_h = (resize_h / shapeSize - 1) * shapeSize;
+
+  if (resize_w % shapeSize == 0)
+    resize_w = resize_w;
+  else if (resize_w / shapeSize < 1 + 1e-5)
+    resize_w = shapeSize;
+  else
+    resize_w = (resize_w / shapeSize - 1) * shapeSize;
+
+  cv::Mat resize_img;
+  cv::resize(img, resize_img, cv::Size(resize_w, resize_h));
+
+  ratio_hw.push_back(static_cast<float>(resize_h) / static_cast<float>(h));
+  ratio_hw.push_back(static_cast<float>(resize_w) / static_cast<float>(w));
+  return resize_img;
+}
+
+void OCR_PPredictor::infer_det(cv::Mat &origin, int max_size_len, std::vector<OCRPredictResult> &ocr_results) {
+  std::vector<float> mean = {0.485f, 0.456f, 0.406f};
+  std::vector<float> scale = {1 / 0.229f, 1 / 0.224f, 1 / 0.225f};
+
+  PredictorInput input = _det_predictor->get_first_input();
+
+  std::vector<float> ratio_hw;
+  cv::Mat input_image = DetResizeImg(origin, max_size_len, ratio_hw);
+  input_image.convertTo(input_image, CV_32FC3, 1 / 255.0f);
+  const float *dimg = reinterpret_cast<const float *>(input_image.data);
+  int input_size = input_image.rows * input_image.cols;
+
+  input.set_dims({1, 3, input_image.rows, input_image.cols});
+
+  neon_mean_scale(dimg, input.get_mutable_float_data(), input_size, mean,
+                  scale);
+  LOGI("ocr cpp det shape %d,%d", input_image.rows,input_image.cols);
+  std::vector<PredictorOutput> results = _det_predictor->infer();
+  PredictorOutput &res = results.at(0);
+  std::vector<std::vector<std::vector<int>>> filtered_box = calc_filtered_boxes(
+          res.get_float_data(), res.get_size(), input_image.rows, input_image.cols, origin);
+  LOGI("ocr cpp det Filter_box size %ld", filtered_box.size());
+
+  for(auto & i : filtered_box){
+    LOGI("ocr cpp box  %d,%d,%d,%d,%d,%d,%d,%d", i[0][0],i[0][1], i[1][0],i[1][1], i[2][0],i[2][1], i[3][0],i[3][1]);
+    OCRPredictResult res;
+    res.points = i;
+    ocr_results.push_back(res);
+  }
+}
+
+void OCR_PPredictor::infer_rec(const cv::Mat &origin_img, int run_cls, OCRPredictResult& ocr_result) {
+  std::vector<float> mean = {0.5f, 0.5f, 0.5f};
+  std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
+  std::vector<int64_t> dims = {1, 3, 0, 0};
+
+  PredictorInput input = _rec_predictor->get_first_input();
+
+  const std::vector<std::vector<int>> &box = ocr_result.points;
+  cv::Mat crop_img;
+  if (!box.empty()) {
+    crop_img = get_rotate_crop_image(origin_img, box);
+  } else {
+    crop_img = origin_img;
+  }
+
+  if (run_cls) {
+    ClsPredictResult cls_res = infer_cls(crop_img);
+    crop_img = cls_res.img;
+    ocr_result.cls_score = cls_res.cls_score;
+    ocr_result.cls_label = cls_res.cls_label;
+  }
+
+
+  float wh_ratio = float(crop_img.cols) / float(crop_img.rows);
+  cv::Mat input_image = crnn_resize_img(crop_img, wh_ratio);
+  input_image.convertTo(input_image, CV_32FC3, 1 / 255.0f);
+  const float *dimg = reinterpret_cast<const float *>(input_image.data);
+  int input_size = input_image.rows * input_image.cols;
+
+  dims[2] = input_image.rows;
+  dims[3] = input_image.cols;
+  input.set_dims(dims);
+
+  neon_mean_scale(dimg, input.get_mutable_float_data(), input_size, mean,
+                  scale);
+
+  std::vector<PredictorOutput> results = _rec_predictor->infer();
+  const float *predict_batch = results.at(0).get_float_data();
+  const std::vector<int64_t> predict_shape = results.at(0).get_shape();
+
+  // ctc decode
+  int argmax_idx;
+  int last_index = 0;
+  float score = 0.f;
+  int count = 0;
+  float max_value = 0.0f;
+
+  for (int n = 0; n < predict_shape[1]; n++) {
+    argmax_idx = int(argmax(&predict_batch[n * predict_shape[2]],
+                            &predict_batch[(n + 1) * predict_shape[2]]));
+    max_value =
+        float(*std::max_element(&predict_batch[n * predict_shape[2]],
+                                &predict_batch[(n + 1) * predict_shape[2]]));
+    if (argmax_idx > 0 && (!(n > 0 && argmax_idx == last_index))) {
+      score += max_value;
+      count += 1;
+      ocr_result.word_index.push_back(argmax_idx);
+    }
+    last_index = argmax_idx;
+  }
+  score /= count;
+  ocr_result.score = score;
+  LOGI("ocr cpp rec word size %ld", count);
+}
+
+ClsPredictResult OCR_PPredictor::infer_cls(const cv::Mat &img, float thresh) {
+  std::vector<float> mean = {0.5f, 0.5f, 0.5f};
+  std::vector<float> scale = {1 / 0.5f, 1 / 0.5f, 1 / 0.5f};
+  std::vector<int64_t> dims = {1, 3, 0, 0};
+
+  PredictorInput input = _cls_predictor->get_first_input();
+
+  cv::Mat input_image = cls_resize_img(img);
+  input_image.convertTo(input_image, CV_32FC3, 1 / 255.0f);
+  const float *dimg = reinterpret_cast<const float *>(input_image.data);
+  int input_size = input_image.rows * input_image.cols;
+
+  dims[2] = input_image.rows;
+  dims[3] = input_image.cols;
+  input.set_dims(dims);
+
+  neon_mean_scale(dimg, input.get_mutable_float_data(), input_size, mean,
+                  scale);
+
+  std::vector<PredictorOutput> results = _cls_predictor->infer();
+
+  const float *scores = results.at(0).get_float_data();
+  float score = 0;
+  int label = 0;
+  for (int64_t i = 0; i < results.at(0).get_size(); i++) {
+    LOGI("ocr cpp cls output scores [%f]", scores[i]);
+    if (scores[i] > score) {
+      score = scores[i];
+      label = i;
+    }
+  }
+  cv::Mat srcimg;
+  img.copyTo(srcimg);
+  if (label % 2 == 1 && score > thresh) {
+    cv::rotate(srcimg, srcimg, 1);
+  }
+  ClsPredictResult res;
+  res.cls_label = label;
+  res.cls_score = score;
+  res.img = srcimg;
+  LOGI("ocr cpp cls word cls %ld, %f", label, score);
+  return res;
+}
+
+std::vector<std::vector<std::vector<int>>>
+OCR_PPredictor::calc_filtered_boxes(const float *pred, int pred_size,
+                                    int output_height, int output_width,
+                                    const cv::Mat &origin) {
+  const double threshold = 0.3;
+  const double maxvalue = 1;
+
+  cv::Mat pred_map = cv::Mat::zeros(output_height, output_width, CV_32F);
+  memcpy(pred_map.data, pred, pred_size * sizeof(float));
+  cv::Mat cbuf_map;
+  pred_map.convertTo(cbuf_map, CV_8UC1);
+
+  cv::Mat bit_map;
+  cv::threshold(cbuf_map, bit_map, threshold, maxvalue, cv::THRESH_BINARY);
+
+  std::vector<std::vector<std::vector<int>>> boxes =
+      boxes_from_bitmap(pred_map, bit_map);
+  float ratio_h = output_height * 1.0f / origin.rows;
+  float ratio_w = output_width * 1.0f / origin.cols;
+  std::vector<std::vector<std::vector<int>>> filter_boxes =
+      filter_tag_det_res(boxes, ratio_h, ratio_w, origin);
+  return filter_boxes;
+}
+
+std::vector<int>
+OCR_PPredictor::postprocess_rec_word_index(const PredictorOutput &res) {
+  const int *rec_idx = res.get_int_data();
+  const std::vector<std::vector<uint64_t>> rec_idx_lod = res.get_lod();
+
+  std::vector<int> pred_idx;
+  for (int n = int(rec_idx_lod[0][0]); n < int(rec_idx_lod[0][1] * 2); n += 2) {
+    pred_idx.emplace_back(rec_idx[n]);
+  }
+  return pred_idx;
+}
+
+float OCR_PPredictor::postprocess_rec_score(const PredictorOutput &res) {
+  const float *predict_batch = res.get_float_data();
+  const std::vector<int64_t> predict_shape = res.get_shape();
+  const std::vector<std::vector<uint64_t>> predict_lod = res.get_lod();
+  int blank = predict_shape[1];
+  float score = 0.f;
+  int count = 0;
+  for (int n = predict_lod[0][0]; n < predict_lod[0][1] - 1; n++) {
+    int argmax_idx = argmax(predict_batch + n * predict_shape[1],
+                            predict_batch + (n + 1) * predict_shape[1]);
+    float max_value = predict_batch[n * predict_shape[1] + argmax_idx];
+    if (blank - 1 - argmax_idx > 1e-5) {
+      score += max_value;
+      count += 1;
+    }
+  }
+  if (count == 0) {
+    LOGE("calc score count 0");
+  } else {
+    score /= count;
+  }
+  LOGI("calc score: %f", score);
+  return score;
+}
+
+NET_TYPE OCR_PPredictor::get_net_flag() const { return NET_OCR; }
+}

libs/paddleocr/src/main/cpp/ocr_ppredictor.h

@@ -0,0 +1,130 @@
+//
+// Created by fujiayi on 2020/7/1.
+//
+
+#pragma once
+
+#include "ppredictor.h"
+#include <opencv2/opencv.hpp>
+#include <paddle_api.h>
+#include <string>
+
+namespace ppredictor {
+
+/**
+ * Config
+ */
+struct OCR_Config {
+    int use_opencl = 0;
+    int thread_num = 4; // Thread num
+  paddle::lite_api::PowerMode mode =
+      paddle::lite_api::LITE_POWER_HIGH; // PaddleLite Mode
+};
+
+/**
+ * PolyGone Result
+ */
+struct OCRPredictResult {
+  std::vector<int> word_index;
+  std::vector<std::vector<int>> points;
+  float score;
+  float cls_score;
+  int cls_label=-1;
+};
+
+struct ClsPredictResult {
+    float cls_score;
+    int cls_label=-1;
+    cv::Mat img;
+};
+/**
+ * OCR there are 2 models
+ * 1. First model（det），select polygones to show where are the texts
+ * 2. crop from the origin images, use these polygones to infer
+ */
+class OCR_PPredictor : public PPredictor_Interface {
+public:
+  OCR_PPredictor(const OCR_Config &config);
+
+  virtual ~OCR_PPredictor() {}
+
+  /**
+   * 初始化二个模型的Predictor
+   * @param det_model_content
+   * @param rec_model_content
+   * @return
+   */
+  int init(const std::string &det_model_content,
+           const std::string &rec_model_content,
+           const std::string &cls_model_content);
+  int init_from_file(const std::string &det_model_path,
+                     const std::string &rec_model_path,
+                     const std::string &cls_model_path);
+  /**
+   * Return OCR result
+   * @param dims
+   * @param input_data
+   * @param input_len
+   * @param net_flag
+   * @param origin
+   * @return
+   */
+  virtual std::vector<OCRPredictResult>
+  infer_ocr(cv::Mat &origin, int max_size_len, int run_det, int run_cls, int run_rec);
+
+  virtual NET_TYPE get_net_flag() const;
+
+private:
+  /**
+   * calcul Polygone from the result image of first model
+   * @param pred
+   * @param output_height
+   * @param output_width
+   * @param origin
+   * @return
+   */
+  std::vector<std::vector<std::vector<int>>>
+  calc_filtered_boxes(const float *pred, int pred_size, int output_height,
+                      int output_width, const cv::Mat &origin);
+
+  void
+  infer_det(cv::Mat &origin, int max_side_len, std::vector<OCRPredictResult>& ocr_results);
+  /**
+   * infer for rec model
+   *
+   * @param boxes
+   * @param origin
+   * @return
+   */
+  void
+  infer_rec(const cv::Mat &origin, int run_cls, OCRPredictResult& ocr_result);
+
+  /**
+  * infer for cls model
+  *
+  * @param boxes
+  * @param origin
+  * @return
+  */
+  ClsPredictResult infer_cls(const cv::Mat &origin, float thresh = 0.9);
+
+  /**
+   * Postprocess or sencod model to extract text
+   * @param res
+   * @return
+   */
+  std::vector<int> postprocess_rec_word_index(const PredictorOutput &res);
+
+  /**
+   * calculate confidence of second model text result
+   * @param res
+   * @return
+   */
+  float postprocess_rec_score(const PredictorOutput &res);
+
+  std::unique_ptr<PPredictor> _det_predictor;
+  std::unique_ptr<PPredictor> _rec_predictor;
+  std::unique_ptr<PPredictor> _cls_predictor;
+  OCR_Config _config;
+};
+}

libs/paddleocr/src/main/cpp/ppredictor.cpp

@@ -0,0 +1,95 @@
+#include "ppredictor.h"
+#include "common.h"
+
+namespace ppredictor {
+PPredictor::PPredictor(int use_opencl, int thread_num, int net_flag,
+                       paddle::lite_api::PowerMode mode)
+    : _use_opencl(use_opencl), _thread_num(thread_num), _net_flag(net_flag), _mode(mode) {}
+
+int PPredictor::init_nb(const std::string &model_content) {
+  paddle::lite_api::MobileConfig config;
+  config.set_model_from_buffer(model_content);
+  return _init(config);
+}
+
+int PPredictor::init_from_file(const std::string &model_content) {
+  paddle::lite_api::MobileConfig config;
+  config.set_model_from_file(model_content);
+  return _init(config);
+}
+
+template <typename ConfigT> int PPredictor::_init(ConfigT &config) {
+  bool is_opencl_backend_valid = paddle::lite_api::IsOpenCLBackendValid(/*check_fp16_valid = false*/);
+  if (is_opencl_backend_valid) {
+    if (_use_opencl != 0) {
+      // Make sure you have write permission of the binary path.
+      // We strongly recommend each model has a unique binary name.
+      const std::string bin_path = "/data/local/tmp/";
+      const std::string bin_name = "lite_opencl_kernel.bin";
+      config.set_opencl_binary_path_name(bin_path, bin_name);
+
+      // opencl tune option
+      // CL_TUNE_NONE: 0
+      // CL_TUNE_RAPID: 1
+      // CL_TUNE_NORMAL: 2
+      // CL_TUNE_EXHAUSTIVE: 3
+      const std::string tuned_path = "/data/local/tmp/";
+      const std::string tuned_name = "lite_opencl_tuned.bin";
+      config.set_opencl_tune(paddle::lite_api::CL_TUNE_NORMAL, tuned_path, tuned_name);
+
+      // opencl precision option
+      // CL_PRECISION_AUTO: 0, first fp16 if valid, default
+      // CL_PRECISION_FP32: 1, force fp32
+      // CL_PRECISION_FP16: 2, force fp16
+      config.set_opencl_precision(paddle::lite_api::CL_PRECISION_FP32);
+      LOGI("ocr cpp device: running on gpu.");
+    }
+  } else {
+    LOGI("ocr cpp device: running on cpu.");
+    // you can give backup cpu nb model instead
+    // config.set_model_from_file(cpu_nb_model_dir);
+  }
+  config.set_threads(_thread_num);
+  config.set_power_mode(_mode);
+  _predictor = paddle::lite_api::CreatePaddlePredictor(config);
+  LOGI("ocr cpp paddle instance created");
+  return RETURN_OK;
+}
+
+PredictorInput PPredictor::get_input(int index) {
+  PredictorInput input{_predictor->GetInput(index), index, _net_flag};
+  _is_input_get = true;
+  return input;
+}
+
+std::vector<PredictorInput> PPredictor::get_inputs(int num) {
+  std::vector<PredictorInput> results;
+  for (int i = 0; i < num; i++) {
+    results.emplace_back(get_input(i));
+  }
+  return results;
+}
+
+PredictorInput PPredictor::get_first_input() { return get_input(0); }
+
+std::vector<PredictorOutput> PPredictor::infer() {
+  LOGI("ocr cpp infer Run start %d", _net_flag);
+  std::vector<PredictorOutput> results;
+  if (!_is_input_get) {
+    return results;
+  }
+  _predictor->Run();
+  LOGI("ocr cpp infer Run end");
+
+  for (int i = 0; i < _predictor->GetOutputNames().size(); i++) {
+    std::unique_ptr<const paddle::lite_api::Tensor> output_tensor =
+        _predictor->GetOutput(i);
+    LOGI("ocr cpp output tensor[%d] size %ld", i, product(output_tensor->shape()));
+    PredictorOutput result{std::move(output_tensor), i, _net_flag};
+    results.emplace_back(std::move(result));
+  }
+  return results;
+}
+
+NET_TYPE PPredictor::get_net_flag() const { return (NET_TYPE)_net_flag; }
+}

libs/paddleocr/src/main/cpp/ppredictor.h

@@ -0,0 +1,64 @@
+#pragma once
+
+#include "paddle_api.h"
+#include "predictor_input.h"
+#include "predictor_output.h"
+
+namespace ppredictor {
+
+/**
+ * PaddleLite Preditor Common Interface
+ */
+class PPredictor_Interface {
+public:
+  virtual ~PPredictor_Interface() {}
+
+  virtual NET_TYPE get_net_flag() const = 0;
+};
+
+/**
+ * Common Predictor
+ */
+class PPredictor : public PPredictor_Interface {
+public:
+  PPredictor(
+      int use_opencl, int thread_num, int net_flag = 0,
+      paddle::lite_api::PowerMode mode = paddle::lite_api::LITE_POWER_HIGH);
+
+  virtual ~PPredictor() {}
+
+  /**
+   * init paddlitelite opt model，nb format ，or use ini_paddle
+   * @param model_content
+   * @return 0
+   */
+  virtual int init_nb(const std::string &model_content);
+
+  virtual int init_from_file(const std::string &model_content);
+
+  std::vector<PredictorOutput> infer();
+
+  std::shared_ptr<paddle::lite_api::PaddlePredictor> get_predictor() {
+    return _predictor;
+  }
+
+  virtual std::vector<PredictorInput> get_inputs(int num);
+
+  virtual PredictorInput get_input(int index);
+
+  virtual PredictorInput get_first_input();
+
+  virtual NET_TYPE get_net_flag() const;
+
+protected:
+  template <typename ConfigT> int _init(ConfigT &config);
+
+private:
+  int _use_opencl;
+  int _thread_num;
+  paddle::lite_api::PowerMode _mode;
+  std::shared_ptr<paddle::lite_api::PaddlePredictor> _predictor;
+  bool _is_input_get = false;
+  int _net_flag;
+};
+}

libs/paddleocr/src/main/cpp/predictor_input.cpp

@@ -0,0 +1,28 @@
+#include "predictor_input.h"
+
+namespace ppredictor {
+
+void PredictorInput::set_dims(std::vector<int64_t> dims) {
+  // yolov3
+  if (_net_flag == 101 && _index == 1) {
+    _tensor->Resize({1, 2});
+    _tensor->mutable_data<int>()[0] = (int)dims.at(2);
+    _tensor->mutable_data<int>()[1] = (int)dims.at(3);
+  } else {
+    _tensor->Resize(dims);
+  }
+  _is_dims_set = true;
+}
+
+float *PredictorInput::get_mutable_float_data() {
+  if (!_is_dims_set) {
+    LOGE("PredictorInput::set_dims is not called");
+  }
+  return _tensor->mutable_data<float>();
+}
+
+void PredictorInput::set_data(const float *input_data, int input_float_len) {
+  float *input_raw_data = get_mutable_float_data();
+  memcpy(input_raw_data, input_data, input_float_len * sizeof(float));
+}
+}

libs/paddleocr/src/main/cpp/predictor_input.h

@@ -0,0 +1,26 @@
+#pragma once
+
+#include "common.h"
+#include <paddle_api.h>
+#include <vector>
+
+namespace ppredictor {
+class PredictorInput {
+public:
+  PredictorInput(std::unique_ptr<paddle::lite_api::Tensor> &&tensor, int index,
+                 int net_flag)
+      : _tensor(std::move(tensor)), _index(index), _net_flag(net_flag) {}
+
+  void set_dims(std::vector<int64_t> dims);
+
+  float *get_mutable_float_data();
+
+  void set_data(const float *input_data, int input_float_len);
+
+private:
+  std::unique_ptr<paddle::lite_api::Tensor> _tensor;
+  bool _is_dims_set = false;
+  int _index;
+  int _net_flag;
+};
+}

libs/paddleocr/src/main/cpp/predictor_output.cpp

@@ -0,0 +1,26 @@
+#include "predictor_output.h"
+namespace ppredictor {
+const float *PredictorOutput::get_float_data() const {
+  return _tensor->data<float>();
+}
+
+const int *PredictorOutput::get_int_data() const {
+  return _tensor->data<int>();
+}
+
+const std::vector<std::vector<uint64_t>> PredictorOutput::get_lod() const {
+  return _tensor->lod();
+}
+
+int64_t PredictorOutput::get_size() const {
+  if (_net_flag == NET_OCR) {
+    return _tensor->shape().at(2) * _tensor->shape().at(3);
+  } else {
+    return product(_tensor->shape());
+  }
+}
+
+const std::vector<int64_t> PredictorOutput::get_shape() const {
+  return _tensor->shape();
+}
+}

libs/paddleocr/src/main/cpp/predictor_output.h

@@ -0,0 +1,31 @@
+#pragma once
+
+#include "common.h"
+#include <paddle_api.h>
+#include <vector>
+
+namespace ppredictor {
+class PredictorOutput {
+public:
+  PredictorOutput() {}
+  PredictorOutput(std::unique_ptr<const paddle::lite_api::Tensor> &&tensor,
+                  int index, int net_flag)
+      : _tensor(std::move(tensor)), _index(index), _net_flag(net_flag) {}
+
+  const float *get_float_data() const;
+  const int *get_int_data() const;
+  int64_t get_size() const;
+  const std::vector<std::vector<uint64_t>> get_lod() const;
+  const std::vector<int64_t> get_shape() const;
+
+  std::vector<float> data;    // return float, or use data_int
+  std::vector<int> data_int;  // several layers return int ，or use data
+  std::vector<int64_t> shape; // PaddleLite output shape
+  std::vector<std::vector<uint64_t>> lod; // PaddleLite output lod
+
+private:
+  std::unique_ptr<const paddle::lite_api::Tensor> _tensor;
+  int _index;
+  int _net_flag;
+};
+}

libs/paddleocr/src/main/cpp/preprocess.cpp

@@ -0,0 +1,82 @@
+#include "preprocess.h"
+#include <android/bitmap.h>
+
+cv::Mat bitmap_to_cv_mat(JNIEnv *env, jobject bitmap) {
+  AndroidBitmapInfo info;
+  int result = AndroidBitmap_getInfo(env, bitmap, &info);
+  if (result != ANDROID_BITMAP_RESULT_SUCCESS) {
+    LOGE("AndroidBitmap_getInfo failed, result: %d", result);
+    return cv::Mat{};
+  }
+  if (info.format != ANDROID_BITMAP_FORMAT_RGBA_8888) {
+    LOGE("Bitmap format is not RGBA_8888 !");
+    return cv::Mat{};
+  }
+  unsigned char *srcData = NULL;
+  AndroidBitmap_lockPixels(env, bitmap, (void **)&srcData);
+  cv::Mat mat = cv::Mat::zeros(info.height, info.width, CV_8UC4);
+  memcpy(mat.data, srcData, info.height * info.width * 4);
+  AndroidBitmap_unlockPixels(env, bitmap);
+  cv::cvtColor(mat, mat, cv::COLOR_RGBA2BGR);
+  /**
+  if (!cv::imwrite("/sdcard/1/copy.jpg", mat)){
+      LOGE("Write image failed " );
+  }
+   */
+
+  return mat;
+}
+
+cv::Mat resize_img(const cv::Mat &img, int height, int width) {
+  if (img.rows == height && img.cols == width) {
+    return img;
+  }
+  cv::Mat new_img;
+  cv::resize(img, new_img, cv::Size(height, width));
+  return new_img;
+}
+
+// fill tensor with mean and scale and trans layout: nhwc -> nchw, neon speed up
+void neon_mean_scale(const float *din, float *dout, int size,
+                     const std::vector<float> &mean,
+                     const std::vector<float> &scale) {
+  if (mean.size() != 3 || scale.size() != 3) {
+    LOGE("[ERROR] mean or scale size must equal to 3");
+    return;
+  }
+
+  float32x4_t vmean0 = vdupq_n_f32(mean[0]);
+  float32x4_t vmean1 = vdupq_n_f32(mean[1]);
+  float32x4_t vmean2 = vdupq_n_f32(mean[2]);
+  float32x4_t vscale0 = vdupq_n_f32(scale[0]);
+  float32x4_t vscale1 = vdupq_n_f32(scale[1]);
+  float32x4_t vscale2 = vdupq_n_f32(scale[2]);
+
+  float *dout_c0 = dout;
+  float *dout_c1 = dout + size;
+  float *dout_c2 = dout + size * 2;
+
+  int i = 0;
+  for (; i < size - 3; i += 4) {
+    float32x4x3_t vin3 = vld3q_f32(din);
+    float32x4_t vsub0 = vsubq_f32(vin3.val[0], vmean0);
+    float32x4_t vsub1 = vsubq_f32(vin3.val[1], vmean1);
+    float32x4_t vsub2 = vsubq_f32(vin3.val[2], vmean2);
+    float32x4_t vs0 = vmulq_f32(vsub0, vscale0);
+    float32x4_t vs1 = vmulq_f32(vsub1, vscale1);
+    float32x4_t vs2 = vmulq_f32(vsub2, vscale2);
+    vst1q_f32(dout_c0, vs0);
+    vst1q_f32(dout_c1, vs1);
+    vst1q_f32(dout_c2, vs2);
+
+    din += 12;
+    dout_c0 += 4;
+    dout_c1 += 4;
+    dout_c2 += 4;
+  }
+  for (; i < size; i++) {
+    *(dout_c0++) = (*(din++) - mean[0]) * scale[0];
+    *(dout_c1++) = (*(din++) - mean[1]) * scale[1];
+    *(dout_c2++) = (*(din++) - mean[2]) * scale[2];
+  }
+}

libs/paddleocr/src/main/cpp/preprocess.h

@@ -0,0 +1,12 @@
+#pragma once
+
+#include "common.h"
+#include <jni.h>
+#include <opencv2/opencv.hpp>
+cv::Mat bitmap_to_cv_mat(JNIEnv *env, jobject bitmap);
+
+cv::Mat resize_img(const cv::Mat &img, int height, int width);
+
+void neon_mean_scale(const float *din, float *dout, int size,
+                     const std::vector<float> &mean,
+                     const std::vector<float> &scale);

libs/paddleocr/src/main/java/com/baidu/paddle/lite/ocr/OCRPredictorNative.java

@@ -0,0 +1,127 @@
+package com.baidu.paddle.lite.ocr;
+
+import android.graphics.Bitmap;
+import android.util.Log;
+
+import org.opencv.android.OpenCVLoader;
+
+import java.util.ArrayList;
+import java.util.concurrent.atomic.AtomicBoolean;
+import java.util.concurrent.locks.ReentrantLock;
+
+/**
+ * @author PaddleOCR
+ * Modified by TonyJiangWJ
+ * @since 2023-08-06
+ */
+public class OCRPredictorNative {
+
+    private static final AtomicBoolean isSOLoaded = new AtomicBoolean();
+    private static final ReentrantLock lock = new ReentrantLock();
+
+    public static void loadLibrary() throws RuntimeException {
+        if (!isSOLoaded.get() && isSOLoaded.compareAndSet(false, true)) {
+            try {
+                // 可能和AJ中的OpenCV冲突，直接初始化一遍
+                OpenCVLoader.initDebug();
+                System.loadLibrary("Native");
+            } catch (Throwable e) {
+                throw new RuntimeException(
+                        "Load libNative.so failed, please check it exists in apk file.", e);
+            }
+        }
+    }
+
+    private Config config;
+
+    private long nativePointer = 0;
+
+    public OCRPredictorNative(Config config) {
+        lock.lock();
+        try {
+            this.config = config;
+            loadLibrary();
+            nativePointer = init(config.detModelFilename, config.recModelFilename, config.clsModelFilename, config.useOpencl,
+                    config.cpuThreadNum, config.cpuPower);
+            Log.i("OCRPredictorNative", "load success " + nativePointer);
+        } finally {
+            lock.unlock();
+        }
+    }
+
+
+    public ArrayList<OcrResultModel> runImage(Bitmap originalImage, int max_size_len, int run_det, int run_cls, int run_rec) {
+        lock.lock();
+        try {
+            Log.i("OCRPredictorNative", "begin to run image");
+            float[] rawResults = forward(nativePointer, originalImage, max_size_len, run_det, run_cls, run_rec);
+            return postprocess(rawResults);
+        } finally {
+            lock.unlock();
+        }
+    }
+
+    public static class Config {
+        public int useOpencl;
+        public int cpuThreadNum;
+        public String cpuPower;
+        public String detModelFilename;
+        public String recModelFilename;
+        public String clsModelFilename;
+
+    }
+
+    public void destroy() {
+        if (nativePointer != 0) {
+            release(nativePointer);
+            nativePointer = 0;
+        }
+    }
+
+    protected native long init(String detModelPath, String recModelPath, String clsModelPath, int useOpencl, int threadNum, String cpuMode);
+
+    protected native float[] forward(long pointer, Bitmap originalImage, int max_size_len, int run_det, int run_cls, int run_rec);
+
+    protected native void release(long pointer);
+
+    private ArrayList<OcrResultModel> postprocess(float[] raw) {
+        ArrayList<OcrResultModel> results = new ArrayList<OcrResultModel>();
+        int begin = 0;
+
+        while (begin < raw.length) {
+            int point_num = Math.round(raw[begin]);
+            int word_num = Math.round(raw[begin + 1]);
+            OcrResultModel model = parse(raw, begin + 2, point_num, word_num);
+            begin += 2 + 1 + point_num * 2 + word_num + 2;
+            results.add(model);
+        }
+
+        return results;
+    }
+
+    private OcrResultModel parse(float[] raw, int begin, int pointNum, int wordNum) {
+        int current = begin;
+        OcrResultModel model = new OcrResultModel();
+        model.setConfidence(raw[current]);
+        current++;
+        for (int i = 0; i < pointNum; i++) {
+            model.addPoints(Math.round(raw[current + i * 2]), Math.round(raw[current + i * 2 + 1]));
+        }
+        current += (pointNum * 2);
+        for (int i = 0; i < wordNum; i++) {
+            int index = Math.round(raw[current + i]);
+            model.addWordIndex(index);
+        }
+        current += wordNum;
+        model.setClsIdx(raw[current]);
+        model.setClsConfidence(raw[current + 1]);
+        Log.i("OCRPredictorNative", "word finished " + wordNum);
+        return model;
+    }
+
+    @Override
+    protected void finalize() throws Throwable {
+        super.finalize();
+        destroy();
+    }
+}

libs/paddleocr/src/main/java/com/baidu/paddle/lite/ocr/OcrResult.java

@@ -0,0 +1,134 @@
+package com.baidu.paddle.lite.ocr;
+
+import android.graphics.Point;
+import android.graphics.Rect;
+
+import java.util.ArrayList;
+import java.util.List;
+
+/**
+ * @author PaddleOCR
+ * Modified by TonyJiangWJ
+ * @since 2023-08-06
+ */
+public class OcrResult implements Comparable<OcrResult> {
+    private String label;
+    private float confidence;
+    private Rect bounds;
+    private final List<OcrResult> elements = new ArrayList<>();
+
+    public OcrResult() {
+    }
+
+    public OcrResult(OcrResultModel resultModel) {
+        this.label = resultModel.getLabel();
+        this.confidence = resultModel.getConfidence();
+        int left = -1, right = -1, top = -1, bottom = -1;
+        for (Point point : resultModel.getPoints()) {
+            if (point.x < left || left == -1) {
+                left = point.x;
+            }
+            if (point.x > right || right == -1) {
+                right = point.x;
+            }
+            if (point.y < top || top == -1) {
+                top = point.y;
+            }
+            if (point.y > bottom || bottom == -1) {
+                bottom = point.y;
+            }
+        }
+        this.bounds = new Rect(left, top, right, bottom);
+    }
+
+    public OcrResult(String label, float confidence, Rect bounds) {
+        this.label = label;
+        this.confidence = confidence;
+        this.bounds = bounds;
+    }
+
+    public String getLabel() {
+        return label;
+    }
+
+    public void setLabel(String label) {
+        this.label = label;
+    }
+
+    public float getConfidence() {
+        return confidence;
+    }
+
+    public void setConfidence(float confidence) {
+        this.confidence = confidence;
+    }
+
+    public Rect getBounds() {
+        return bounds;
+    }
+
+    public void setBounds(Rect bounds) {
+        this.bounds = bounds;
+    }
+
+    public RectLocation getLocation() {
+        return new RectLocation(bounds);
+    }
+
+    public String getWords() {
+        return label.trim().replace("\r", "");
+    }
+
+    public List<OcrResult> getElements() {
+        return this.elements;
+    }
+
+    public void addElements(OcrResult element) {
+        this.elements.add(element);
+    }
+
+    @Override
+    public int compareTo(OcrResult o) {
+        // 上下差距小于二分之一的高度 判定为同一行
+        int deviation = Math.max(this.bounds.height(), o.bounds.height()) / 2;
+        // 通过垂直中心点的距离判定
+        if (Math.abs((this.bounds.top + this.bounds.bottom) / 2 - (o.bounds.top + o.bounds.bottom) / 2) < deviation) {
+            return this.bounds.left - o.bounds.left;
+        } else {
+            return this.bounds.bottom - o.bounds.bottom;
+        }
+    }
+
+    @Override
+    public String toString() {
+        return "OcrResult{" + "label='" + label + '\'' +
+                ", confidence=" + confidence +
+                ", bounds=" + bounds +
+                ", elements=" + elements +
+                '}';
+    }
+
+    public static class RectLocation {
+        public int left;
+        public int top;
+        public int width;
+        public int height;
+
+        public RectLocation() {
+        }
+
+        public RectLocation(int left, int top, int width, int height) {
+            this.left = left;
+            this.top = top;
+            this.width = width;
+            this.height = height;
+        }
+
+        public RectLocation(Rect rect) {
+            left = rect.left;
+            top = rect.top;
+            width = rect.right - rect.left;
+            height = rect.bottom - rect.top;
+        }
+    }
+}

libs/paddleocr/src/main/java/com/baidu/paddle/lite/ocr/OcrResultModel.java

@@ -0,0 +1,97 @@
+package com.baidu.paddle.lite.ocr;
+
+import android.graphics.Point;
+
+import java.util.ArrayList;
+import java.util.List;
+
+/**
+ * @author PaddleOCR
+ * Modified by TonyJiangWJ
+ * @since 2023-08-06
+ */
+public class OcrResultModel {
+    private final List<Point> points;
+    private final List<Integer> wordIndex;
+    private String label;
+    private float confidence;
+    private float clsIdx;
+    private String clsLabel;
+    private float clsConfidence;
+
+    public OcrResultModel() {
+        super();
+        points = new ArrayList<>();
+        wordIndex = new ArrayList<>();
+    }
+
+    public void addPoints(int x, int y) {
+        Point point = new Point(x, y);
+        points.add(point);
+    }
+
+    public void addWordIndex(int index) {
+        wordIndex.add(index);
+    }
+
+    public List<Point> getPoints() {
+        return points;
+    }
+
+    public List<Integer> getWordIndex() {
+        return wordIndex;
+    }
+
+    public String getLabel() {
+        return label;
+    }
+
+    public void setLabel(String label) {
+        this.label = label;
+    }
+
+    public float getConfidence() {
+        return confidence;
+    }
+
+    public void setConfidence(float confidence) {
+        this.confidence = confidence;
+    }
+
+    public float getClsIdx() {
+        return clsIdx;
+    }
+
+    public void setClsIdx(float idx) {
+        this.clsIdx = idx;
+    }
+
+    public String getClsLabel() {
+        return clsLabel;
+    }
+
+    public void setClsLabel(String label) {
+        this.clsLabel = label;
+    }
+
+    public float getClsConfidence() {
+        return clsConfidence;
+    }
+
+    public void setClsConfidence(float confidence) {
+        this.clsConfidence = confidence;
+    }
+
+    @Override
+    public String toString() {
+        return "OcrResultModel{" +
+                "points=" + points +
+                ", wordIndex=" + wordIndex +
+                ", label='" + label + '\'' +
+                ", confidence=" + confidence +
+                ", clsIdx=" + clsIdx +
+                ", clsLabel='" + clsLabel + '\'' +
+                ", clsConfidence=" + clsConfidence +
+                '}';
+    }
+}

libs/paddleocr/src/main/java/com/baidu/paddle/lite/ocr/Predictor.java

@@ -0,0 +1,338 @@
+package com.baidu.paddle.lite.ocr;
+
+import android.content.Context;
+import android.graphics.Bitmap;
+import android.graphics.BitmapFactory;
+import android.preference.PreferenceManager;
+import android.util.Base64;
+import android.util.Log;
+
+import java.io.File;
+import java.io.InputStream;
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.Collections;
+import java.util.Date;
+import java.util.List;
+import java.util.Vector;
+
+/**
+ * @author PaddleOCR
+ * Modified by TonyJiangWJ
+ * @since 2023-08-06
+ */
+public class Predictor {
+    private static final String TAG = Predictor.class.getSimpleName();
+    public boolean isLoaded = false;
+    public int warmupIterNum = 1;
+    public int inferIterNum = 1;
+    public int cpuThreadNum = 4;
+    public String cpuPowerMode = "LITE_POWER_HIGH";
+    public String modelPath = "";
+    public String modelName = "";
+    protected OCRPredictorNative paddlePredictor = null;
+    protected float inferenceTime = 0;
+    // Only for object detection
+    protected Vector<String> wordLabels = new Vector<>();
+    protected int detLongSize = 960;
+    public float scoreThreshold = 0.1f;
+    protected Bitmap inputImage = null;
+    protected float preprocessTime = 0;
+    /**
+     * 自定义开关
+     */
+    public boolean useSlim = true;
+    public boolean useOpencl = false;
+    public boolean checkModelLoaded = true;
+    public boolean runCls = false;
+    public boolean runDet = false;
+    public boolean runRec = true;
+
+    private static final String CHECK_IMG_BASE64 = "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";
+    private static final Bitmap checkingBitmap = BitmapFactory.decodeByteArray(Base64.decode(CHECK_IMG_BASE64, Base64.DEFAULT), 0, Base64.decode(CHECK_IMG_BASE64, Base64.DEFAULT).length);
+    /**
+     * 检测模型
+     */
+    public String detModelFilename = "det_opt.nb";
+    /**
+     * 识别模型
+     */
+    public String recModelFilename = "rec_opt.nb";
+    /**
+     * 文本方向检测模型
+     */
+    public String clsModelFilename = "cls_opt.nb";
+
+    private final String defaultLabelPath = "labels/ppocr_keys_v1.txt";
+    private final String defaultModelPath = "models/ocr_v3_for_cpu";
+    /**
+     * slim模型 目前使用的是2.10版的opt工具转换的2.11版本不能正常使用
+     */
+    private final String defaultModelPathSlim = "models/ocr_v3_for_cpu(slim)";
+
+    private int retryTime = 1;
+
+    public Predictor() {
+    }
+
+    public boolean init(Context appCtx) {
+        return this.init(appCtx, defaultModelPath, defaultLabelPath);
+    }
+
+    public boolean init(Context appCtx, boolean useSlim) {
+        if (this.isLoaded && this.useSlim == useSlim) {
+            return true;
+        }
+        this.useSlim = useSlim;
+        if (useSlim) {
+            return this.init(appCtx, defaultModelPathSlim, defaultLabelPath);
+        } else {
+            return this.init(appCtx, defaultModelPath, defaultLabelPath);
+        }
+    }
+
+    public boolean init(Context appCtx, String modelPath, String labelPath) {
+        isLoaded = loadModel(appCtx, modelPath, cpuThreadNum, cpuPowerMode);
+        if (!isLoaded) {
+            return false;
+        }
+        isLoaded = loadLabel(appCtx, labelPath);
+        if (!checkModelLoadedSuccess()) {
+            init(appCtx, modelPath, labelPath);
+        }
+        return isLoaded;
+    }
+
+    /**
+     * 初始化模型后通过识别预设图片校验是否初始化成功
+     * 曲线救国 深层的失败原因需要后续排查
+     *
+     * @return
+     */
+    private boolean checkModelLoadedSuccess() {
+        if (!checkModelLoaded) {
+            return true;
+        }
+        if (!isLoaded) {
+            return false;
+        }
+        List<OcrResult> results = runOcr(checkingBitmap);
+        StringBuilder sb = new StringBuilder();
+        for (OcrResult result : results) {
+            sb.append(result.getLabel());
+        }
+        boolean check = sb.toString().contains("测试");
+        Log.d(TAG, "第" + retryTime + "次 校验是否初始化成功: " + check + " 识别结果：" + sb);
+        boolean result = check || retryTime++ >= 5;
+        if (!check && retryTime >= 5) {
+            Log.e(TAG, "初始化模型失败");
+        }
+        return result;
+    }
+
+    public boolean init(Context appCtx, String modelPath, String labelPath, int cpuThreadNum, String cpuPowerMode) {
+        isLoaded = loadModel(appCtx, modelPath, cpuThreadNum, cpuPowerMode);
+        if (!isLoaded) {
+            return false;
+        }
+        isLoaded = loadLabel(appCtx, labelPath);
+        return isLoaded;
+    }
+
+    public boolean init(Context appCtx, String modelPath, String labelPath, int cpuThreadNum, String cpuPowerMode,
+                        int detLongSize, float scoreThreshold) {
+        boolean isLoaded = init(appCtx, modelPath, labelPath, cpuThreadNum, cpuPowerMode);
+        if (!isLoaded) {
+            return false;
+        }
+        this.detLongSize = detLongSize;
+        this.scoreThreshold = scoreThreshold;
+        return true;
+    }
+
+    protected boolean loadModel(Context appCtx, String modelPath, int cpuThreadNum, String cpuPowerMode) {
+        // Release model if exists
+        releaseModel();
+
+        // Load model
+        if (modelPath.isEmpty()) {
+            return false;
+        }
+        String realPath = modelPath;
+        if (modelPath.charAt(0) != '/') {
+            // Read model files from custom path if the first character of mode path is '/'
+            // otherwise copy model to cache from assets
+            realPath = appCtx.getCacheDir() + "/" + modelPath;
+            // region add by TonyJiangWJ
+            String key = "PADDLE_MODEL_LOADED";
+            // 进行了模型更新 需要强制覆盖旧模型
+            boolean loaded = PreferenceManager.getDefaultSharedPreferences(appCtx).getBoolean(key, false);
+            if (loaded) {
+                // 没有必要每次都复制
+                Utils.copyDirectoryFromAssetsIfNeeded(appCtx, modelPath, realPath);
+            } else {
+                Utils.copyDirectoryFromAssets(appCtx, modelPath, realPath);
+                PreferenceManager.getDefaultSharedPreferences(appCtx).edit().putBoolean(key, true).apply();
+            }
+            // endregion
+        }
+
+        OCRPredictorNative.Config config = new OCRPredictorNative.Config();
+        // 是否使用GPU
+        config.useOpencl = useOpencl ? 1 : 0;
+        config.cpuThreadNum = cpuThreadNum;
+        config.detModelFilename = realPath + File.separator + detModelFilename;
+        config.recModelFilename = realPath + File.separator + recModelFilename;
+        config.clsModelFilename = realPath + File.separator + clsModelFilename;
+        Log.i("Predictor", "model path" + config.detModelFilename + " ; " + config.recModelFilename + ";" + config.clsModelFilename);
+        config.cpuPower = cpuPowerMode;
+        paddlePredictor = new OCRPredictorNative(config);
+
+        this.cpuThreadNum = cpuThreadNum;
+        this.cpuPowerMode = cpuPowerMode;
+        this.modelPath = realPath;
+        this.modelName = realPath.substring(realPath.lastIndexOf("/") + 1);
+        return true;
+    }
+
+    public void releaseModel() {
+        if (paddlePredictor != null) {
+            paddlePredictor.destroy();
+            paddlePredictor = null;
+        }
+        isLoaded = false;
+        modelPath = "";
+        modelName = "";
+    }
+
+    protected boolean loadLabel(Context appCtx, String labelPath) {
+        wordLabels.clear();
+        wordLabels.add("black");
+        // Load word labels from file
+        try {
+            InputStream assetsInputStream = appCtx.getAssets().open(labelPath);
+            int available = assetsInputStream.available();
+            byte[] lines = new byte[available];
+            if (assetsInputStream.read(lines) <= 0) {
+                Log.e(TAG, "读取label失败");
+                return false;
+            }
+            assetsInputStream.close();
+            String words = new String(lines);
+            String[] contents = words.split("\n");
+            wordLabels.addAll(Arrays.asList(contents));
+            wordLabels.add(" ");
+            Log.i(TAG, "Word label size: " + wordLabels.size());
+        } catch (Exception e) {
+            Log.e(TAG, e.getMessage());
+            return false;
+        }
+        return true;
+    }
+
+
+    public List<OcrResult> runOcr(Bitmap inputImage) {
+        if (inputImage == null || !isLoaded()) {
+            return Collections.emptyList();
+        }
+        // 检测、分类、识别
+        int run_det = runDet ? 1 : 0, run_cls = runCls ? 1 : 0, run_rec = runRec ? 1 : 0;
+        // Warm up
+        for (int i = 0; i < warmupIterNum; i++) {
+            paddlePredictor.runImage(inputImage, detLongSize, run_det, run_cls, run_rec);
+        }
+        warmupIterNum = 0; // do not need warm
+        // Run inference
+        Date start = new Date();
+        ArrayList<OcrResultModel> results = paddlePredictor.runImage(inputImage, detLongSize, run_det, run_cls, run_rec);
+        Date end = new Date();
+        inferenceTime = (end.getTime() - start.getTime()) / (float) inferIterNum;
+
+        postProcess(results);
+        Log.i(TAG, "[stat] Preprocess Time: " + preprocessTime
+                + " ; Inference Time: " + inferenceTime + " ;Box Size " + results.size());
+        List<OcrResult> ocrResults = new ArrayList<>();
+        for (OcrResultModel resultModel : results) {
+            Log.d(TAG, "recognize: " + resultModel.toString());
+            if (resultModel.getConfidence() >= scoreThreshold) {
+                ocrResults.add(new OcrResult(resultModel));
+            }
+        }
+        Collections.sort(ocrResults);
+        return ocrResults;
+    }
+
+
+    public boolean isLoaded() {
+        return paddlePredictor != null && isLoaded;
+    }
+
+    public String modelPath() {
+        return modelPath;
+    }
+
+    public String modelName() {
+        return modelName;
+    }
+
+    public int cpuThreadNum() {
+        return cpuThreadNum;
+    }
+
+    public String cpuPowerMode() {
+        return cpuPowerMode;
+    }
+
+    public float inferenceTime() {
+        return inferenceTime;
+    }
+
+    public Bitmap inputImage() {
+        return inputImage;
+    }
+
+    public float preprocessTime() {
+        return preprocessTime;
+    }
+
+    public String getDefaultLabelPath() {
+        return defaultLabelPath;
+    }
+
+    public String getDefaultModelPath() {
+        return defaultModelPath;
+    }
+
+    public String getDefaultModelPathSlim() {
+        return defaultModelPathSlim;
+    }
+
+    public boolean isUseSlim() {
+        return useSlim;
+    }
+
+    public void setInputImage(Bitmap image) {
+        if (image == null) {
+            return;
+        }
+        this.inputImage = image.copy(Bitmap.Config.ARGB_8888, true);
+    }
+
+    private void postProcess(ArrayList<OcrResultModel> results) {
+        for (OcrResultModel r : results) {
+            StringBuilder word = new StringBuilder();
+            for (int index : r.getWordIndex()) {
+                if (index >= 0 && index < wordLabels.size()) {
+                    word.append(wordLabels.get(index).replace("\r", ""));
+                } else {
+                    Log.e(TAG, "Word index is not in label list:" + index);
+                    word.append(" ");
+                }
+            }
+            r.setLabel(word.toString());
+            r.setClsLabel(r.getClsIdx() == 1 ? "180" : "0");
+        }
+    }
+
+
+}

libs/paddleocr/src/main/java/com/baidu/paddle/lite/ocr/Utils.java

@@ -0,0 +1,199 @@
+package com.baidu.paddle.lite.ocr;
+
+import android.content.Context;
+import android.graphics.Bitmap;
+import android.graphics.Matrix;
+import android.media.ExifInterface;
+import android.os.Environment;
+import android.util.Log;
+
+import java.io.BufferedInputStream;
+import java.io.BufferedOutputStream;
+import java.io.File;
+import java.io.FileNotFoundException;
+import java.io.FileOutputStream;
+import java.io.IOException;
+import java.io.InputStream;
+import java.io.OutputStream;
+
+/**
+ * @author PaddleOCR
+ * Modified by TonyJiangWJ
+ * @since 2023-08-06
+ */
+public class Utils {
+    private static final String TAG = Utils.class.getSimpleName();
+
+    public static void copyFileFromAssets(Context appCtx, String srcPath, String dstPath) {
+        if (srcPath.isEmpty() || dstPath.isEmpty()) {
+            return;
+        }
+        InputStream is = null;
+        OutputStream os = null;
+        try {
+            is = new BufferedInputStream(appCtx.getAssets().open(srcPath));
+            os = new BufferedOutputStream(new FileOutputStream(new File(dstPath)));
+            byte[] buffer = new byte[1024];
+            int length = 0;
+            while ((length = is.read(buffer)) != -1) {
+                os.write(buffer, 0, length);
+            }
+        } catch (FileNotFoundException e) {
+            e.printStackTrace();
+        } catch (IOException e) {
+            e.printStackTrace();
+        } finally {
+            try {
+                os.close();
+                is.close();
+            } catch (IOException e) {
+                e.printStackTrace();
+            }
+        }
+    }
+
+    public static void copyDirectoryFromAssets(Context appCtx, String srcDir, String dstDir) {
+        if (srcDir.isEmpty() || dstDir.isEmpty()) {
+            return;
+        }
+        try {
+            if (!new File(dstDir).exists()) {
+                new File(dstDir).mkdirs();
+            }
+            for (String fileName : appCtx.getAssets().list(srcDir)) {
+                String srcSubPath = srcDir + File.separator + fileName;
+                String dstSubPath = dstDir + File.separator + fileName;
+                if (new File(srcSubPath).isDirectory()) {
+                    copyDirectoryFromAssets(appCtx, srcSubPath, dstSubPath);
+                } else {
+                    Log.d(TAG, "复制资源文件: " + srcSubPath + " => " + dstSubPath);
+                    copyFileFromAssets(appCtx, srcSubPath, dstSubPath);
+                }
+            }
+        } catch (Exception e) {
+            e.printStackTrace();
+        }
+    }
+
+    public static void copyDirectoryFromAssetsIfNeeded(Context appCtx, String srcDir, String dstDir) {
+        if (srcDir.isEmpty() || dstDir.isEmpty()) {
+            return;
+        }
+        try {
+            if (!new File(dstDir).exists()) {
+                new File(dstDir).mkdirs();
+            }
+            for (String fileName : appCtx.getAssets().list(srcDir)) {
+                String srcSubPath = srcDir + File.separator + fileName;
+                String dstSubPath = dstDir + File.separator + fileName;
+                if (new File(srcSubPath).isDirectory()) {
+                    copyDirectoryFromAssetsIfNeeded(appCtx, srcSubPath, dstSubPath);
+                } else {
+                    if (new File(dstSubPath).exists()) {
+                        return;
+                    }
+                    Log.d(TAG, "复制资源文件: " + srcSubPath + " => " + dstSubPath);
+                    copyFileFromAssets(appCtx, srcSubPath, dstSubPath);
+                }
+            }
+        } catch (Exception e) {
+            e.printStackTrace();
+        }
+    }
+
+
+    public static float[] parseFloatsFromString(String string, String delimiter) {
+        String[] pieces = string.trim().toLowerCase().split(delimiter);
+        float[] floats = new float[pieces.length];
+        for (int i = 0; i < pieces.length; i++) {
+            floats[i] = Float.parseFloat(pieces[i].trim());
+        }
+        return floats;
+    }
+
+    public static long[] parseLongsFromString(String string, String delimiter) {
+        String[] pieces = string.trim().toLowerCase().split(delimiter);
+        long[] longs = new long[pieces.length];
+        for (int i = 0; i < pieces.length; i++) {
+            longs[i] = Long.parseLong(pieces[i].trim());
+        }
+        return longs;
+    }
+
+    public static String getSDCardDirectory() {
+        return Environment.getExternalStorageDirectory().getAbsolutePath();
+    }
+
+    public static boolean isSupportedNPU() {
+        return false;
+        // String hardware = android.os.Build.HARDWARE;
+        // return hardware.equalsIgnoreCase("kirin810") || hardware.equalsIgnoreCase("kirin990");
+    }
+
+    public static Bitmap resizeWithStep(Bitmap bitmap, int maxLength, int step) {
+        int width = bitmap.getWidth();
+        int height = bitmap.getHeight();
+        int maxWH = Math.max(width, height);
+        float ratio = 1;
+        int newWidth = width;
+        int newHeight = height;
+        if (maxWH > maxLength) {
+            ratio = maxLength * 1.0f / maxWH;
+            newWidth = (int) Math.floor(ratio * width);
+            newHeight = (int) Math.floor(ratio * height);
+        }
+
+        newWidth = newWidth - newWidth % step;
+        if (newWidth == 0) {
+            newWidth = step;
+        }
+        newHeight = newHeight - newHeight % step;
+        if (newHeight == 0) {
+            newHeight = step;
+        }
+        return Bitmap.createScaledBitmap(bitmap, newWidth, newHeight, true);
+    }
+
+    public static Bitmap rotateBitmap(Bitmap bitmap, int orientation) {
+
+        Matrix matrix = new Matrix();
+        switch (orientation) {
+            case ExifInterface.ORIENTATION_NORMAL:
+                return bitmap;
+            case ExifInterface.ORIENTATION_FLIP_HORIZONTAL:
+                matrix.setScale(-1, 1);
+                break;
+            case ExifInterface.ORIENTATION_ROTATE_180:
+                matrix.setRotate(180);
+                break;
+            case ExifInterface.ORIENTATION_FLIP_VERTICAL:
+                matrix.setRotate(180);
+                matrix.postScale(-1, 1);
+                break;
+            case ExifInterface.ORIENTATION_TRANSPOSE:
+                matrix.setRotate(90);
+                matrix.postScale(-1, 1);
+                break;
+            case ExifInterface.ORIENTATION_ROTATE_90:
+                matrix.setRotate(90);
+                break;
+            case ExifInterface.ORIENTATION_TRANSVERSE:
+                matrix.setRotate(-90);
+                matrix.postScale(-1, 1);
+                break;
+            case ExifInterface.ORIENTATION_ROTATE_270:
+                matrix.setRotate(-90);
+                break;
+            default:
+                return bitmap;
+        }
+        try {
+            Bitmap bmRotated = Bitmap.createBitmap(bitmap, 0, 0, bitmap.getWidth(), bitmap.getHeight(), matrix, true);
+            bitmap.recycle();
+            return bmRotated;
+        } catch (OutOfMemoryError e) {
+            e.printStackTrace();
+            return null;
+        }
+    }
+}

settings.gradle.kts

@@ -10,6 +10,7 @@ include(
     ":libs:jackpal.androidterm.emulatorview-1.0.42",
     ":libs:jackpal.androidterm.libtermexec-1.0",
     ":libs:org.opencv-4.5.5",
+    ":libs:paddleocr",
 
     ":libs:android-spackle-9.0.0",
     ":libs:android-assertion-9.0.0",