添加PaddleOCR功能

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);