概述
OpenCV4.0深度神经网络模块，支持openface模型的导入，提取人脸的128特征向量，进行相似度比对，实现人脸识别。Openface模型的详细信息看这里
http://www.cv-foundation.org/openaccess/content_cvpr_2015/app/1A_089.pdf
主要原理是基于2015年CVPR的FaceNet网络的论文，去年的时候写过一篇文章介绍过它，想要了解详细信息的点击这里查看即可
主要思路
首先使用OpenCV4.0 DNN模块支持的人脸检测模型，实现对图像或者视频的人脸检测，然后对得到的人脸区域通过openface的预训练模型提取128个特征向量值，基于余弦相似度进行特征值比对，实现人脸识别。完整的流程可以图示如下：
image
余弦相似公式与解释：
image
代码实现步骤
01
加载网络
需要先加载人脸检测与openface人脸识别网络模型，代码实现如下：
String modelDesc = "D:/projects/opencv_tutorial/data/models/resnet/deploy.prototxt"; String modelBinary = "D:/projects/opencv_tutorial/data/models/resnet/res10_300x300_ssd_iter_140000.caffemodel"; String facemodel = "D:/projects/opencv_tutorial/data/models/face_detector/openface.nn4.small2.v1.t7"; // 初始化网络 Net net = readNetFromCaffe(modelDesc, modelBinary); Net netRecogn = readNetFromTorch(facemodel);
这两个模型的下载地址如下：
https://github.com/gloomyfish1998/opencv_tutorial/tree/master/data/models/face_detector
02
设置计算后台
OpenCV支持不同的计算后台，这里我们采用OpenVINO作为计算后台，可以实现加速计算，代码如下：
// 设置计算后台Net netRecogn = readNetFromTorch(facemodel);net.setPreferableBackend(DNN_BACKEND_INFERENCE_ENGINE);net.setPreferableTarget(DNN_TARGET_CPU);netRecogn.setPreferableBackend(DNN_BACKEND_INFERENCE_ENGINE);netRecogn.setPreferableTarget(DNN_TARGET_CPU);// load face datavector<vector<float>> face_data;vector<string> labels;vector<string> faces;glob("D:/my_faces/zhigang", faces);for (auto fn : faces) { vector<float> fv; Mat sample = imread(fn); recognize_face(sample, netRecogn, fv); face_data.push_back(fv); printf("file name : %s\n", fn.c_str()); labels.push_back("zhigang");}faces.clear();glob("D:/my_faces/balvin", faces);for (auto fn : faces) { vector<float> fv; Mat sample = imread(fn); recognize_face(sample, netRecogn, fv); face_data.push_back(fv); printf("file name : %s\n", fn.c_str()); labels.push_back("balvin");}if (net.empty() || netRecogn.empty()){ printf("could not load net...\n"); return -1;}
03
人脸检测
通过人脸检测网络实现人脸检测，代码实现如下：
// 输入数据调整Mat inputBlob = blobFromImage(frame, inScaleFactor, Size(inWidth, inHeight), meanVal, false, false);net.setInput(inputBlob, "data");// 人脸检测Mat detection = net.forward("detection_out");vector<double> layersTimings;double freq = getTickFrequency() / 1000;double time = net.getPerfProfile(layersTimings) / freq;Mat detectionMat(detection.size[2], detection.size[3], CV_32F, detection.ptr<float>());
04
人脸比对
把实时检测得到ROI区域与预先加载的人脸样本进行比较，找到距离最小的，如果小于阈值T，即为识别输出结果，解析人脸检测并实现人脸识别的代码如下：
for (int i = 0; i < detectionMat.rows; i++){ // 置信度 0～1之间 float confidence = detectionMat.at<float>(i, 2); if (confidence > confidenceThreshold) { int xLeftBottom = static_cast<int>(detectionMat.at<float>(i, 3) * frame.cols); int yLeftBottom = static_cast<int>(detectionMat.at<float>(i, 4) * frame.rows); int xRightTop = static_cast<int>(detectionMat.at<float>(i, 5) * frame.cols); int yRightTop = static_cast<int>(detectionMat.at<float>(i, 6) * frame.rows); Rect object((int)xLeftBottom, (int)yLeftBottom, (int)(xRightTop - xLeftBottom), (int)(yRightTop - yLeftBottom)); if (object.width < 5 || object.height < 5) { continue; } // 截取人脸ROI区域 Mat roi = frame(object); // 人脸比对，发现相似度最高的 vector<float> curr_fv; recognize_face(roi, netRecogn, curr_fv); float minDist = 10; int index = -1; for (int i = 0; i < face_data.size(); i++) { float dist = compare(curr_fv, face_data[i]); if (minDist > dist) { minDist = dist; index = i; } } // 阈值与显示识别结果 printf("index : %d, dist: %.2f \n", index, minDist); if (index >= 0 && minDist < 0.30) { putText(frame, labels[index].c_str(), Point(xLeftBottom, yLeftBottom-20), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(255, 0, 255)); } rectangle(frame, object, Scalar(0, 255, 0)); ss.str(""); ss << confidence; String conf(ss.str()); String label = "Face: " + conf; int baseLine = 0; Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine); rectangle(frame, Rect(Point(xLeftBottom, yLeftBottom - labelSize.height), Size(labelSize.width, labelSize.height + baseLine)), Scalar(255, 255, 255), FILLED); putText(frame, label, Point(xLeftBottom, yLeftBottom), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 0, 0)); }}
余弦相似比较
float compare(vector<float> &fv1, vector<float> fv2) { // 计算余弦相似, 0 ~ 1 距离，距离越小越相似， // 0表示夹角为0°，1表示夹角为90° float dot = 0; float sum2 = 0; float sum3 = 0; for (int i = 0; i < fv1.size(); i++) { dot += fv1[i] * fv2[i]; sum2 += pow(fv1[i], 2); sum3 += pow(fv2[i], 2); } float norm = sqrt(sum2)*sqrt(sum3); float similarity = dot / norm; float dist = acos(similarity) / CV_PI; return dist;}
运行效果
image
网友评论