基于DNN的人脸检测与识别

目录

• 引言
• 模型
• 代码
• 解释
  • 人脸识别
• 参考文献
• 致谢

上一篇教程： 高级API：TextDetectionModel 和 TextRecognitionModel
下一篇教程： PyTorch分类模型的转换和使用OpenCV Python启动

原作者	王成锐，冯元涛
兼容性	OpenCV >= 4.5.4

引言

在本节中，我们介绍用于人脸检测的cv::FaceDetectorYN类和用于人脸识别的cv::FaceRecognizerSF类。

模型

此模块预训练并需要两个模型（ONNX格式）

• 人脸检测:
  • 大小：338KB
  • 在WIDER Face Val数据集上的结果：0.830（简单）、0.824（中等）、0.708（困难）
• 人脸识别
  • 大小：36.9MB
  • 结果

数据库	准确率	阈值 (normL2)	阈值 (cosine)
LFW	99.60%	1.128	0.363
CALFW	93.95%	1.149	0.340
CPLFW	91.05%	1.204	0.275
AgeDB-30	94.90%	1.202	0.277
CFP-FP	94.80%	1.253	0.212

代码

C++

• 可下载代码： 点击 这里
• 代码概览

  #include <opencv2/dnn.hpp>
  #include <opencv2/imgproc.hpp>
  #include <opencv2/highgui.hpp>
  #include <opencv2/objdetect.hpp>
   
  #include <iostream>
   
  using namespace cv;
  using namespace std;
   
  static
  void visualize(Mat& input, int frame, Mat& faces, double fps, int thickness = 2)
  {
  std::string fpsString = cv::format("FPS : %.2f", (float)fps);
      if (frame >= 0)
  cout << "帧 " << frame << ", ";
  cout << "FPS: " << fpsString << endl;
      for (int i = 0; i < faces.rows; i++)
      {
          // 打印结果
  cout << "人脸 " << i
  << ", 左上角坐标: (" << faces.at<float>(i, 0) << ", " << faces.at<float>(i, 1) << "), "
  << "框宽度: " << faces.at<float>(i, 2) << ", 框高度: " << faces.at<float>(i, 3) << ", "
  << "得分: " << cv::format("%.2f", faces.at<float>(i, 14))
  << endl;
   
          // 绘制边界框
  rectangle(input, Rect2i(int(faces.at<float>(i, 0)), int(faces.at<float>(i, 1)), int(faces.at<float>(i, 2)), int(faces.at<float>(i, 3))), Scalar(0, 255, 0), thickness);
          // 绘制关键点
  circle(input, Point2i(int(faces.at<float>(i, 4)), int(faces.at<float>(i, 5))), 2, Scalar(255, 0, 0), thickness);
  circle(input, Point2i(int(faces.at<float>(i, 6)), int(faces.at<float>(i, 7))), 2, Scalar(0, 0, 255), thickness);
  circle(input, Point2i(int(faces.at<float>(i, 8)), int(faces.at<float>(i, 9))), 2, Scalar(0, 255, 0), thickness);
  circle(input, Point2i(int(faces.at<float>(i, 10)), int(faces.at<float>(i, 11))), 2, Scalar(255, 0, 255), thickness);
  circle(input, Point2i(int(faces.at<float>(i, 12)), int(faces.at<float>(i, 13))), 2, Scalar(0, 255, 255), thickness);
      }
  putText(input, fpsString, Point(0, 15), FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 255, 0), 2);
  }
   
  int main(int argc, char** argv)
  {
      CommandLineParser parser(argc, argv,
          “{help h | | 打印此消息}”
          “{image1 i1 | | 输入图像1的路径。检测视频流时可省略}”
          “{image2 i2 | | 输入图像2的路径。当给出image1和image2参数时，程序将尝试在两张图像上查找人脸并运行人脸识别算法}”
          “{video v | 0 | 输入视频的路径}”
          “{scale sc | 1.0 | 用于调整输入视频帧大小的缩放因子}”
          “{fd_model fd | face_detection_yunet_2021dec.onnx| 模型路径。从 https://github.com/opencv/opencv_zoo/tree/master/models/face_detection_yunet 下载 yunet.onnx}”
          “{fr_model fr | face_recognition_sface_2021dec.onnx | 人脸识别模型的路径。从 https://github.com/opencv/opencv_zoo/tree/master/models/face_recognition_sface 下载模型}”
          “{score_threshold | 0.9 | 过滤掉得分 < score_threshold 的人脸}”
          “{nms_threshold | 0.3 | 抑制iou >= nms_threshold 的边界框}”
          “{top_k | 5000 | 在NMS之前保留top_k个边界框}”
          “{save s | false | 设置为true以保存结果。使用摄像头时此标志无效}”
      );
      if (parser.has("help"))
      {
  parser.printMessage();
          return 0; 0;
      }
   
      String fd_modelPath = parser.get<String>("fd_model");
      String fr_modelPath = parser.get<String>("fr_model");
   
      float scoreThreshold = parser.get<float>("score_threshold");
      float nmsThreshold = parser.get<float>("nms_threshold");
      int topK = parser.get<int>("top_k");
   
      bool save = parser.get<bool>("save");
      float scale = parser.get<float>("scale");
   
      double cosine_similar_thresh = 0.363;
      double l2norm_similar_thresh = 1.128;
   
      // 初始化 FaceDetectorYN
      Ptr<FaceDetectorYN> detector = FaceDetectorYN::create(fd_modelPath, "", Size(320, 320), scoreThreshold, nmsThreshold, topK);
   
      TickMeter tm;
   
      // 如果输入是图像
      if (parser.has("image1"))
      {
          String input1 = parser.get<String>("image1");
          Mat image1 = imread(samples::findFile(input1));
          if (image1.empty())
          {
  std::cerr << "无法读取图像: " << input1 << std::endl;
              return 0; 2;
          }
   
          int imageWidth = int(image1.cols * scale);
          int imageHeight = int(image1.rows * scale);
          resize(image1, image1, Size(imageWidth, imageHeight));
  tm.start();
   
          // 设置推理前的输入大小
  detector->setInputSize(image1.size());
   
          Mat faces1;
  detector->detect(image1, faces1);
          if (faces1.rows < 1)
          {
  std::cerr << "无法在 " << input1 << " 中找到人脸" << std::endl;
              return 0; 1;
          }
   
  tm.stop();
          // 在输入图像上绘制结果
  visualize(image1, -1, faces1, tm.getFPS());
   
          // 如果save为true，则保存结果
          if (save)
          {
  cout << "正在保存 result.jpg...\n";
              imwrite("result.jpg", image1);
          }
   
          // 可视化结果
          imshow("image1", image1);
          pollKey(); // 处理UI事件以显示内容
   
          if (parser.has("image2"))
          {
              String input2 = parser.get<String>("image2");
              Mat image2 = imread(samples::findFile(input2));
              if (image2.empty())
              {
  std::cerr << "无法读取image2: " << input2 << std::endl;
                  return 0; 2;
              }
   
  tm.reset();
  tm.start();
  detector->setInputSize(image2.size());
   
              Mat faces2;
  detector->detect(image2, faces2);
              if (faces2.rows < 1)
              {
  std::cerr << "无法在 " << input2 << " 中找到人脸" << std::endl;
                  return 0; 1;
              }
  tm.stop();
  visualize(image2, -1, faces2, tm.getFPS());
              if (save)
              {
  cout << "正在保存 result2.jpg...\n";
                  imwrite("result2.jpg", image2);
              }
              imshow("image2", image2);
              pollKey();();
   
              // 初始化 FaceRecognizerSF
              Ptr<FaceRecognizerSF> faceRecognizer = FaceRecognizerSF::create(fr_modelPath, "");
   
   
              // 通过检测到的人脸的第一张脸进行人脸图像的对齐和裁剪。
              Mat aligned_face1, aligned_face2;
  faceRecognizer->alignCrop(image1, faces1.row(0), aligned_face1);
  faceRecognizer->alignCrop(image2, faces2.row(0), aligned_face2);
   
              // 使用对齐后的 aligned_face 执行特征提取
              Mat feature1, feature2;
  faceRecognizer->feature(aligned_face1, feature1);
  feature1 = feature1.clone();
  faceRecognizer->feature(aligned_face2, feature2);
  feature2 = feature2.clone();
   
              double cos_score = faceRecognizer->match(feature1, feature2, FaceRecognizerSF::DisType::FR_COSINE);
              double L2_score = faceRecognizer->match(feature1, feature2, FaceRecognizerSF::DisType::FR_NORM_L2);
   
              if (cos_score >= cosine_similar_thresh)
              {
  std::cout << "他们具有相同的身份；";
              }
              else
              {
  std::cout << "他们具有不同的身份；";
              }
  std::cout << " 余弦相似度：" << cos_score << "，阈值：" << cosine_similar_thresh << ". (值越高，相似度越高，最大值为 1.0)\n";
   
              if (L2_score <= l2norm_similar_thresh)
              {
  std::cout << "他们具有相同的身份；";
              }
              else
              {
  std::cout << "他们具有相同的身份。";
              }
  std::cout << " L2范数距离：" << L2_score << "，阈值：" << l2norm_similar_thresh << ". (值越低，相似度越高，最小值为 0.0)\n";
          }
  cout << "按任意键退出..." << endl;
          waitKey(0);
      }
      else
      {
          int frameWidth, frameHeight;
          VideoCapture capture;
  std::string video = parser.get<string>("video");
          if (video.size() == 1 && isdigit(video[0]))
  capture.open(parser.get<int>("video"));
          else
  capture.open(samples::findFileOrKeep(video)); // 保持 GStreamer 管道
          if (capture.isOpened())
          {
  frameWidth = int(capture.get(CAP_PROP_FRAME_WIDTH) * scale);
  frameHeight = int(capture.get(CAP_PROP_FRAME_HEIGHT) * scale);
  cout << "视频 " << video
  << ": 宽度=" << frameWidth
  << ", 高度=" << frameHeight
  << endl;
          }
          else
          {
  cout << "无法初始化视频捕获：" << video << "\n";
              return 0; 1;
          }
   
  detector->setInputSize(Size(frameWidth, frameHeight));
   
  cout << "按空格键保存帧，按其他键退出..." << endl;
          int nFrame = 0;
          for (;;)
          {
              // 获取帧
              Mat frame;
              if (!capture.read(frame))
              {
  cerr << "无法抓取帧！停止\n";
                  break;
              }
   
              resize(frame, frame, Size(frameWidth, frameHeight));
   
              // 推理
              Mat faces;
  tm.start();
  detector->detect(frame, faces);
  tm.stop();
   
              Mat result = frame.clone();
              // 在输入图像上绘制结果
  visualize(result, nFrame, faces, tm.getFPS());
   
              // 可视化结果
              imshow("实时画面", result);
   
              int key = waitKey(1);
              bool saveFrame = save;
              if (key == ' ')
              {
  saveFrame = true;
  key = 0; // 已处理
              }
   
              if (saveFrame)
              {
  std::string frame_name = cv::format("frame_%05d.png", nFrame);
  std::string result_name = cv::format("result_%05d.jpg", nFrame);
  cout << "正在保存 '" << frame_name << "' 和 '" << result_name << "' ...\n";
                  imwrite(frame_name, frame);
                  imwrite(result_name, result);
              }
   
  ++nFrame;
   
              if (key > 0)
                  break;
          }
  cout << "已处理 " << nFrame << " 帧" << endl;
      }
  cout << "完成。" << endl;
      return 0; 0;
  }

Python

• 可下载代码：点击这里
• 代码概览

  import argparse
   
  import numpy as np
  import cv2 as cv
   
  def str2bool(v)
      if v.lower() in ['on', 'yes', 'true', 'y', 't']
          return True
      elif v.lower() in ['off', 'no', 'false', 'n', 'f']
          return False
      else:
          raise NotImplementedError
   
  parser = argparse.ArgumentParser()
  parser.add_argument('--image1', '-i1', type=str, help='输入图像1的路径。省略此参数则使用默认摄像头进行检测。')
  parser.add_argument('--image2', '-i2', type=str, help='输入图像2的路径。当给出image1和image2参数时，程序将尝试在这两张图像上查找人脸并运行人脸识别算法。')
  parser.add_argument('--video', '-v', type=str, help='输入视频的路径。')
  parser.add_argument('--scale', '-sc', type=float, default=1.0, help='用于调整输入视频帧大小的缩放因子。')
  parser.add_argument('--face_detection_model', '-fd', type=str, default='face_detection_yunet_2021dec.onnx', help='人脸检测模型的路径。请从 https://github.com/opencv/opencv_zoo/tree/master/models/face_detection_yunet 下载模型。')
  parser.add_argument('--face_recognition_model', '-fr', type=str, default='face_recognition_sface_2021dec.onnx', help='人脸识别模型的路径。请从 https://github.com/opencv/opencv_zoo/tree/master/models/face_recognition_sface 下载模型。')
  parser.add_argument('--score_threshold', type=float, default=0.9, help='过滤掉得分 < score_threshold 的人脸。')
  parser.add_argument('--nms_threshold', type=float, default=0.3, help='抑制 iou >= nms_threshold 的边界框。')
  parser.add_argument('--top_k', type=int, default=5000, help='在 NMS 之前保留 top_k 个边界框。')
  parser.add_argument('--save', '-s', type=str2bool, default=False, help='设置为 true 以保存结果。使用摄像头时此标志无效。')
  args = parser.parse_args()
   
  def visualize(input, faces, fps, thickness=2)
      if faces[1] is not None
          for idx, face in enumerate(faces[1])
  print('人脸 {}, 左上角坐标: ({:.0f}, {:.0f}), 框宽度: {:.0f}, 框高度 {:.0f}, 得分: {:.2f}'.format(idx, face[0], face[1], face[2], face[3], face[-1]))
   
  coords = face[:-1].astype(np.int32)
              cv.rectangle(input, (coords[0], coords[1]), (coords[0]+coords[2], coords[1]+coords[3]), (0, 255, 0), thickness)
              cv.circle(input, (coords[4], coords[5]), 2, (255, 0, 0), thickness)
              cv.circle(input, (coords[6], coords[7]), 2, (0, 0, 255), thickness)
              cv.circle(input, (coords[8], coords[9]), 2, (0, 255, 0), thickness)
              cv.circle(input, (coords[10], coords[11]), 2, (255, 0, 255), thickness)
              cv.circle(input, (coords[12], coords[13]), 2, (0, 255, 255), thickness)
      cv.putText(input, 'FPS: {:.2f}'.format(fps), (1, 16), cv.FONT_HERSHEY_SIMPLEX, 0.5, (0, 255, 0), 2)
   
  if __name__ == '__main__'
   
      
  detector = cv.FaceDetectorYN.create(
  args.face_detection_model,
          "",
          (320, 320),
  args.score_threshold,
  args.nms_threshold,
  args.top_k
      )
      
   
  tm = cv.TickMeter()
   
      # 如果输入是图像
      if args.image1 is not None
  img1 = cv.imread(cv.samples.findFile(args.image1))
  img1Width = int(img1.shape[1]*args.scale)
  img1Height = int(img1.shape[0]*args.scale)
   
  img1 = cv.resize(img1, (img1Width, img1Height))
  tm.start()
   
          
  detector.setInputSize((img1Width, img1Height))
   
  faces1 = detector.detect(img1)
          
   
  tm.stop()
          assert faces1[1] is not None, '无法在 {} 中找到人脸'.format(args.image1)
   
          # 在输入图像上绘制结果
  visualize(img1, faces1, tm.getFPS())
   
          # 如果 save 为 true，则保存结果
          if args.save
  print('结果已保存到 result.jpg\n')
              cv.imwrite('result.jpg', img1)
   
          # 在新窗口中显示结果
          cv.imshow("image1", img1)
   
          如果 args.image2 不为 空
  img2 = cv.imread(cv.samples.findFile(args.image2))
   
  tm.reset()
  tm.start()
  detector.setInputSize((img2.shape[1], img2.shape[0]))
  faces2 = detector.detect(img2)
  tm.stop()
              断言 faces2[1] 不为 空, '在{}中找不到人脸'.format(args.image2)
  visualize(img2, faces2, tm.getFPS())
              cv.imshow("image2", img2)
   
              
  recognizer = cv.FaceRecognizerSF.create(
  args.face_recognition_model,"")
              
   
              
  face1_align = recognizer.alignCrop(img1, faces1[1][0])
  face2_align = recognizer.alignCrop(img2, faces2[1][0])
   
              # 提取特征
  face1_feature = recognizer.feature(face1_align)
  face2_feature = recognizer.feature(face2_align)
              
   
  cosine_similarity_threshold = 0.363
  l2_similarity_threshold = 1.128
   
              
  cosine_score = recognizer.match(face1_feature, face2_feature, cv.FaceRecognizerSF_FR_COSINE)
  l2_score = recognizer.match(face1_feature, face2_feature, cv.FaceRecognizerSF_FR_NORM_L2)
              
   
  msg = '不同的身份'
              如果 cosine_score >= cosine_similarity_threshold
  msg = '相同的身份'
  print('他们具有{}。余弦相似度：{}，阈值：{} (值越高相似度越高，最大值1.0)。'.format(msg, cosine_score, cosine_similarity_threshold))
   
  msg = '不同的身份'
              如果 l2_score <= l2_similarity_threshold
  msg = '相同的身份'
  print('他们具有{}。NormL2距离：{}，阈值：{} (值越低相似度越高，最小值0.0)。'.format(msg, l2_score, l2_similarity_threshold))
          cv.waitKey(0)
      否则: # 省略输入以调用默认摄像头
          如果 args.video 不为 空
  deviceId = args.video
          else:
  deviceId = 0
  cap = cv.VideoCapture(deviceId)
  frameWidth = int(cap.get(cv.CAP_PROP_FRAME_WIDTH)*args.scale)
  frameHeight = int(cap.get(cv.CAP_PROP_FRAME_HEIGHT)*args.scale)
  detector.setInputSize([frameWidth, frameHeight])
   
          当 cv.waitKey(1) < 0
  hasFrame, frame = cap.read()
              如果 不 hasFrame
  print('没有抓取到帧！')
                  break
   
  frame = cv.resize(frame, (frameWidth, frameHeight))
   
              # 推理
  tm.start()
  faces = detector.detect(frame) # faces 是一个元组
  tm.stop()
   
              # 在输入图像上绘制结果
  visualize(frame, faces, tm.getFPS())
   
              # 可视化结果
              cv.imshow('实时', frame)
      cv.destroyAllWindows()

解释

C++

    // 初始化 FaceDetectorYN
    Ptr<FaceDetectorYN> detector = FaceDetectorYN::create(fd_modelPath, "", Size(320, 320), scoreThreshold, nmsThreshold, topK);

        // 设置推理前的输入大小
detector->setInputSize(image1.size());
 
        Mat faces1;
detector->detect(image1, faces1);
        if (faces1.rows < 1)
        {
std::cerr << "无法在 " << input1 << " 中找到人脸" << std::endl;
            return 0; 1;
        }

Python

detector = cv.FaceDetectorYN.create(
args.face_detection_model,
        "",
        (320, 320),
args.score_threshold,
args.nms_threshold,
args.top_k
    )

detector.setInputSize((img1Width, img1Height))
 
faces1 = detector.detect(img1)

检测输出faces是一个二维CV_32F类型数组，其行是检测到的人脸实例，列是人脸的位置和5个面部特征点。每一行的格式如下：

x1, y1, w, h, x_re, y_re, x_le, y_le, x_nt, y_nt, x_rcm, y_rcm, x_lcm, y_lcm

其中x1, y1, w, h是人脸边界框的左上角坐标、宽度和高度，{x, y}_{re, le, nt, rcm, lcm}分别代表右眼、左眼、鼻尖、嘴巴右角和左角的坐标。

人脸识别

在人脸检测之后，运行下面的代码从面部图像中提取人脸特征。

C++

            // 初始化 FaceRecognizerSF
            Ptr<FaceRecognizerSF> faceRecognizer = FaceRecognizerSF::create(fr_modelPath, "");

            // 通过检测到的人脸的第一张脸进行人脸图像的对齐和裁剪。
            Mat aligned_face1, aligned_face2;
faceRecognizer->alignCrop(image1, faces1.row(0), aligned_face1);
faceRecognizer->alignCrop(image2, faces2.row(0), aligned_face2);
 
            // 使用对齐后的 aligned_face 执行特征提取
            Mat feature1, feature2;
faceRecognizer->feature(aligned_face1, feature1);
feature1 = feature1.clone();
faceRecognizer->feature(aligned_face2, feature2);
feature2 = feature2.clone();

Python

recognizer = cv.FaceRecognizerSF.create(
args.face_recognition_model,"")

face1_align = recognizer.alignCrop(img1, faces1[1][0])
face2_align = recognizer.alignCrop(img2, faces2[1][0])
 
            # 提取特征
face1_feature = recognizer.feature(face1_align)
face2_feature = recognizer.feature(face2_align)

获得两张面部图像的人脸特征feature1和feature2后，运行下面的代码计算这两张人脸的身份差异。

C++

            double cos_score = faceRecognizer->match(feature1, feature2, FaceRecognizerSF::DisType::FR_COSINE);
            double L2_score = faceRecognizer->match(feature1, feature2, FaceRecognizerSF::DisType::FR_NORM_L2);

Python

cosine_score = recognizer.match(face1_feature, face2_feature, cv.FaceRecognizerSF_FR_COSINE)
l2_score = recognizer.match(face1_feature, face2_feature, cv.FaceRecognizerSF_FR_NORM_L2)

例如，如果余弦距离大于或等于0.363，或者NormL2距离小于或等于1.128，则这两张人脸具有相同的身份。

参考

• https://github.com/ShiqiYu/libfacedetection
• https://github.com/ShiqiYu/libfacedetection.train
• https://github.com/zhongyy/SFace

致谢

感谢Yu Shiqi教授和冯元涛的模型训练和提供。

感谢邓老师、钟同学和王同学的模型训练和提供。