CPU實時幀率超過100+，嵌入式端首選的人臉檢測庫推薦
- 2020 年 3 月 10 日
- 筆記
開源人臉庫調用嘗試
基於卷積神經網路的人臉檢測庫，實現了硬體指令加速，純C++ SDK介面，不依賴第三方庫與介面，可以獨立使用，同時也提供了模型文件下載，之前的版本是caffe模型，最新版本支援pytorch。現已開源！支援最小檢測人臉10×10大小
OpenCV DNN可以直接調用訓練好的caffe模型文件，實現實時人臉檢測，演示程式碼如下：
1#include <opencv2/opencv.hpp> 2#include <opencv2/dnn.hpp> 3#include <iostream> 4 5using namespace cv; 6using namespace cv::dnn; 7using namespace std; 8 9const size_t inWidth = 300; 10const size_t inHeight = 300; 11const double inScaleFactor = 0.007843; 12const Scalar meanVal(104.0, 117, 123.0); 13const float confidence = 0.5; 14 15int main(int argc, char** argv) { 16 string model = "D:/projects/models/yufacedetectnet-open-v2.caffemodel"; 17 string config = "D:/projects/models/yufacedetectnet-open-v2.prototxt"; 18 // read network 19 Net net = readNetFromCaffe(config, model); 20 VideoCapture cap(0); 21 Mat frame; 22 while (true) { 23 bool ret = cap.read(frame); 24 if (!ret) { 25 break; 26 } 27 int64 start = getTickCount(); 28 Mat input_data = blobFromImage(frame, inScaleFactor, Size(320, 240), meanVal, false, false); 29 net.setInput(input_data); 30 31 // 人臉檢測 32 Mat detection = net.forward(); 33 Mat detectionMat(detection.size[2], detection.size[3], CV_32F, detection.ptr<float>()); 34 35 // 推斷時間 36 vector<double> layersTimings; 37 double freq = getTickFrequency() / 1000; 38 double time = net.getPerfProfile(layersTimings) / freq; 39 40 ostringstream ss; 41 for (int i = 0; i < detectionMat.rows; i++) 42 { 43 // 置信度 0～1之間 44 float score = detectionMat.at<float>(i, 2); 45 if (score > confidence) 46 { 47 int xLeftBottom = static_cast<int>(detectionMat.at<float>(i, 3) * frame.cols); 48 int yLeftBottom = static_cast<int>(detectionMat.at<float>(i, 4) * frame.rows); 49 int xRightTop = static_cast<int>(detectionMat.at<float>(i, 5) * frame.cols); 50 int yRightTop = static_cast<int>(detectionMat.at<float>(i, 6) * frame.rows); 51 52 Rect object((int)xLeftBottom, (int)yLeftBottom, 53 (int)(xRightTop - xLeftBottom), 54 (int)(yRightTop - yLeftBottom)); 55 56 rectangle(frame, object, Scalar(0, 255, 0)); 57 58 ss << score; 59 String conf(ss.str()); 60 String label = "Face: " + conf; 61 int baseLine = 0; 62 Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine); 63 rectangle(frame, Rect(Point(xLeftBottom, yLeftBottom - labelSize.height), 64 Size(labelSize.width, labelSize.height + baseLine)), 65 Scalar(255, 255, 255), FILLED); 66 putText(frame, label, Point(xLeftBottom, yLeftBottom), 67 FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 0, 0)); 68 } 69 } 70 float fps = getTickFrequency() / (getTickCount() - start); 71 ss.str(""); 72 ss << "FPS: " << fps << " ; inference time: " << time << " ms"; 73 putText(frame, ss.str(), Point(20, 20), 0, 0.75, Scalar(0, 0, 255), 2, 8); 74 imshow("dnn_face_detection", frame); 75 if (waitKey(1) >= 0) break; 76 } 77 waitKey(0); 78 return 0; 79}
運行效果如下：
Github地址
https://github.com/ShiqiYu/libfacedetection
