99 lines
5.6 KiB
TypeScript
99 lines
5.6 KiB
TypeScript
import { highestRisk, riskFromFaceSimilarity, type PortraitClearanceResult, type PortraitImageQuality, type PortraitLocalPrecheck, type PortraitRiskLevel, type PortraitVisionComparison } from "./contracts.js";
|
|
import { colorHistogramCorrelation, structuralSimilarity, type DecodedPortraitImage } from "./image-metrics.js";
|
|
import type { PortraitFaceAnalysis } from "./face-engine.js";
|
|
|
|
export function buildLocalPrecheck(
|
|
imageA: DecodedPortraitImage,
|
|
imageB: DecodedPortraitImage,
|
|
facesA?: PortraitFaceAnalysis,
|
|
facesB?: PortraitFaceAnalysis,
|
|
extraIssues: string[] = [],
|
|
): PortraitLocalPrecheck {
|
|
const reliabilityIssues = [...extraIssues];
|
|
const selectedFaceA = toFaceContract(facesA?.selectedFace);
|
|
const selectedFaceB = toFaceContract(facesB?.selectedFace);
|
|
if (imageA.quality.grade === "poor") reliabilityIssues.push("图片 A 质量过低，模糊、过暗、过曝或对比度不足会导致人脸特征不可靠。");
|
|
if (imageB.quality.grade === "poor") reliabilityIssues.push("图片 B 质量过低，模糊、过暗、过曝或对比度不足会导致人脸特征不可靠。");
|
|
if ((facesA?.faces.length ?? 0) > 1) reliabilityIssues.push("图片 A 检测到多张人脸，未指定目标对象，无法可靠判断。");
|
|
if ((facesB?.faces.length ?? 0) > 1) reliabilityIssues.push("图片 B 检测到多张人脸，未指定目标对象，无法可靠判断。");
|
|
if (facesA?.selectedFace && facesA.selectedFace.detScore < 0.65) reliabilityIssues.push(`图片 A 人脸检测置信度较低（${facesA.selectedFace.detScore.toFixed(4)}），特征提取不稳定。`);
|
|
if (facesB?.selectedFace && facesB.selectedFace.detScore < 0.65) reliabilityIssues.push(`图片 B 人脸检测置信度较低（${facesB.selectedFace.detScore.toFixed(4)}），特征提取不稳定。`);
|
|
if (facesA?.selectedFace && facesA.selectedFace.areaRatio < 0.02) reliabilityIssues.push(`图片 A 人脸区域占比过小（${facesA.selectedFace.areaRatio.toFixed(4)}），本地特征比对不可靠。`);
|
|
if (facesB?.selectedFace && facesB.selectedFace.areaRatio < 0.02) reliabilityIssues.push(`图片 B 人脸区域占比过小（${facesB.selectedFace.areaRatio.toFixed(4)}），本地特征比对不可靠。`);
|
|
return {
|
|
qualityA: imageA.quality,
|
|
qualityB: imageB.quality,
|
|
facesA: facesA?.faces.length ?? 0,
|
|
facesB: facesB?.faces.length ?? 0,
|
|
...(selectedFaceA ? { selectedFaceA } : {}),
|
|
...(selectedFaceB ? { selectedFaceB } : {}),
|
|
...(facesA?.embedding && facesB?.embedding ? { faceSimilarity: round(cosine(facesA.embedding, facesB.embedding), 4) } : {}),
|
|
ssim: structuralSimilarity({ width: imageA.width, height: imageA.height, values: imageA.gray }, { width: imageB.width, height: imageB.height, values: imageB.gray }),
|
|
colorHistogramCorrelation: colorHistogramCorrelation(imageA.rgb, imageB.rgb),
|
|
canExtractEmbedding: Boolean(facesA?.embedding && facesB?.embedding),
|
|
reliabilityIssues: unique(reliabilityIssues),
|
|
};
|
|
}
|
|
|
|
export function localRisk(precheck: PortraitLocalPrecheck): PortraitRiskLevel {
|
|
if (!precheck.canExtractEmbedding || precheck.reliabilityIssues.length > 0 || precheck.facesA !== 1 || precheck.facesB !== 1) return "unable_to_determine";
|
|
return riskFromFaceSimilarity(precheck.faceSimilarity);
|
|
}
|
|
|
|
export function applyStylizedRiskFloor(comparison: PortraitVisionComparison): PortraitVisionComparison {
|
|
if (comparison.analysisPath !== "B" || comparison.featureComparison.distinctive_features.similarity !== "high") return comparison;
|
|
const risk = riskAtLeast(comparison.riskLevel, "medium");
|
|
if (risk === comparison.riskLevel) return comparison;
|
|
return {
|
|
...comparison,
|
|
riskLevel: risk,
|
|
basis: [...comparison.basis, "风格化路径中标志性特征被判断为高度相似，综合风险下限提升至中风险，建议人工复核。"],
|
|
manualReviewRecommended: true,
|
|
};
|
|
}
|
|
|
|
export function summarizeRisk(pairs: Array<{ riskLevel: PortraitRiskLevel }>) {
|
|
const riskCounts: Partial<Record<PortraitRiskLevel, number>> = {};
|
|
for (const pair of pairs) riskCounts[pair.riskLevel] = (riskCounts[pair.riskLevel] ?? 0) + 1;
|
|
return { highestRisk: highestRisk(pairs.map((pair) => pair.riskLevel)), riskCounts };
|
|
}
|
|
|
|
export function buildReliabilityIssues(quality: PortraitImageQuality, faces: number) {
|
|
const issues: string[] = [];
|
|
if (quality.grade === "poor") issues.push("图片质量过低，无法可靠提取肖像特征。");
|
|
if (faces === 0) issues.push("未检测到人脸。");
|
|
if (faces > 1) issues.push("检测到多张人脸，未指定目标对象。");
|
|
return issues;
|
|
}
|
|
|
|
function toFaceContract(face: PortraitFaceAnalysis["selectedFace"]) {
|
|
if (!face) return undefined;
|
|
return { bbox: face.bbox, detScore: face.detScore, areaRatio: face.areaRatio } as const;
|
|
}
|
|
|
|
function cosine(a: ArrayLike<number>, b: ArrayLike<number>) {
|
|
let dot = 0;
|
|
let left = 0;
|
|
let right = 0;
|
|
for (let index = 0; index < a.length; index += 1) {
|
|
dot += Number(a[index]) * Number(b[index]);
|
|
left += Number(a[index]) ** 2;
|
|
right += Number(b[index]) ** 2;
|
|
}
|
|
return dot / Math.max(Number.EPSILON, Math.sqrt(left * right));
|
|
}
|
|
|
|
function riskAtLeast(current: PortraitRiskLevel, floor: PortraitRiskLevel) {
|
|
const order: PortraitRiskLevel[] = ["unable_to_determine", "low", "low_to_medium", "medium", "high"];
|
|
return order[Math.max(order.indexOf(current), order.indexOf(floor))]!;
|
|
}
|
|
|
|
function unique(values: string[]) {
|
|
return [...new Set(values)];
|
|
}
|
|
|
|
function round(value: number, digits: number) {
|
|
const factor = 10 ** digits;
|
|
return Math.round(value * factor) / factor;
|
|
}
|