利用OpenCV和深度學習實現人臉檢測

今天偷點兒懶，就沒有及時整理最新的paper，還請各位看官諒解。這裡整理了一份前段時間做的小demo，實現獻醜了

本文基於OpenCV3.3.1或以上版本（如OpenCV3.4）、DNN模塊和face_detector示例實現簡單、實時的人臉檢測。

往期回顧

[計算機視覺] 入門學習資料

[計算機視覺論文速遞] 2018-03-20

[計算機視覺論文速遞] 2018-03-18

註：

[1]：主要參考Face detection with OpenCV and deep learning這個英文教程，並作部分修改。

[2]：親測OpenCV3.3.0及以下版本，並沒有face_detector示例，且不支持face_detector。為了避免折騰，還是建議使用OpenCV3.3.1及以上（如OpenCV3.4）。

1 face_detector簡介

face_detector示例連結：

https://github.com/opencv/opencv/tree/master/samples/dnn/face_detector

當電腦配置好OpenCV3.3.1或以上版本時，在opencv\samples\dnn也可以找到face_detector示例文件夾，如下圖所示：

使用OpenCV的DNN模塊以及Caffe模型，必須要有.prototxt和.caffemodel兩種文件。但face_detector文件夾中，只有.prototxt一類文件，即缺少訓練好的.caffemodel。.prototxt和.caffemodel的作用如下：

The .prototxt file(s) which define the model architecture (i.e., the layers themselves)

The .caffemodel file which contains the weights for the actual layers

face_detector文件分析：

本教程直接使用訓練好的.caffemodel來進行人臉檢測，即只需要.caffemodel和deploy.prototxt兩個文件。如果想要使用自己的數據集來訓練網絡，請參考"how_to_train_face_detector.txt"。

2 ResNet-10和SSD簡介

本教程屬於實戰篇，故不深入介紹算法內容，若對ResNet和SSD感興趣的同學，可以參考下述連結進行學習

[1]ResNet paper：https://arxiv.org/abs/1512.03385

[2]ResNet in Caffe：https://github.com/soeaver/caffe-model/tree/master/cls/resnet

[3]SSD paper：https://arxiv.org/abs/1512.02325

[4]SSD in Caffe：https://github.com/weiliu89/caffe/tree/ssd

3 .caffemodel下載

res10_300x300_ssd_iter_140000.caffemodel下載連結：https://anonfile.com/W7rdG4d0b1/face_detector.rar

4 C++版本代碼

4.1 圖像中的人臉檢測

對於OpenCV3.4版本，可直接使用opencv-3.4.1\samples\dnn文件夾中的resnet_ssd_face.cpp；

對於OpenCV3.3.1版本，可參考下述代碼：

face_detector_image.cpp

 1
 2
 3
 4
 5
 6
 7#include <iostream>
 8#include <opencv2/opencv.hpp>
 9#include <opencv2/dnn.hpp>
10
11using namespace std;
12using namespace cv;
13using namespace cv::dnn;
14
15
16
17const size_t inWidth = 300;
18const size_t inHeight = 300;
19const double inScaleFactor = 1.0;
20const Scalar meanVal(104.0, 177.0, 123.0);
21
22
23
24int main(int argc, char** argv)
25{
26    
27    Mat img;
28    
29#if 0
30
31    if (argc < 2)
32    {
33        cerr<< "please input "<< endl;
34        cerr << "[Format]face_detector_img.exe image.jpg"<< endl;
35        return -1;
36    }
37
38    img = imread(argv[1]);
39
40#else
41    
42    img = imread("iron_chic.jpg");
43#endif
44
45
46
47    
48    float min_confidence = 0.5;
49
50    String modelConfiguration = "face_detector/deploy.prototxt";
51
52    String modelBinary = "face_detector/res10_300x300_ssd_iter_140000.caffemodel";
53
54    dnn::Net net = readNetFromCaffe(modelConfiguration, modelBinary);
55
56
57
58    if (net.empty())
59    {
60        cerr << "Can't load network by using the following files: " << endl;
61        cerr << "prototxt:   " << modelConfiguration << endl;
62        cerr << "caffemodel: " << modelBinary << endl;
63        cerr << "Models are available here:" << endl;
64        cerr << "<OPENCV_SRC_DIR>/samples/dnn/face_detector" << endl;
65        cerr << "or here:" << endl;
66        cerr << "https://github.com/opencv/opencv/tree/master/samples/dnn/face_detector" << endl;
67        exit(-1);
68    }
69
70
71
72    
73    Mat inputBlob = blobFromImage(img, inScaleFactor, Size(inWidth, inHeight), meanVal, false, false);
74    net.setInput(inputBlob, "data");    
75    Mat detection = net.forward("detection_out");    
76
77    
78
79    vector<double> layersTimings;
80    double freq = getTickFrequency() / 1000;
81    double time = net.getPerfProfile(layersTimings) / freq;
82
83    Mat detectionMat(detection.size[2], detection.size[3], CV_32F, detection.ptr<float>());
84    ostringstream ss;
85    ss << "FPS: " << 1000 / time << " ; time: " << time << "ms" << endl;
86
87    putText(img, ss.str(), Point(20,20), 0, 0.5, Scalar(0, 0, 255));
88
89    float confidenceThreshold = min_confidence;
90    for (int i = 0; i < detectionMat.rows; ++i)
91    {
92        
93        float confidence = detectionMat.at<float>(i, 2);
94
95        if (confidence > confidenceThreshold)
96        {
97            int xLeftBottom = static_cast<int>(detectionMat.at<float>(i, 3) * img.cols);
98            int yLeftBottom = static_cast<int>(detectionMat.at<float>(i, 4) * img.rows);
99            int xRightTop = static_cast<int>(detectionMat.at<float>(i, 5) * img.cols);
100            int yRightTop = static_cast<int>(detectionMat.at<float>(i, 6) * img.rows);
101            Rect object((int)xLeftBottom, (int)yLeftBottom, (int (xRightTop - xLeftBottom),
102            (int)(yRightTop - yLeftBottom));
103            rectangle(img, object, Scalar(0, 255, 0));
104            ss.str("");
105            ss << confidence;
106            String conf(ss.str());
107            String label = "Face: " + conf;
108            int baseLine = 0;
109            Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);
110            rectangle(img, Rect(Point(xLeftBottom, yLeftBottom-labelSize.height),
111            Size(labelSize.width, labelSize.height + baseLine)),
112            Scalar(255, 255, 255), CV_FILLED);
113            putText(img, label, Point(xLeftBottom, yLeftBottom),
114            FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 0, 0));
115
116        }
117    }
118
119    namedWindow("Face Detection", WINDOW_NORMAL);
120    imshow("Face Detection", img);
121    waitKey(0);
122
123return 0;
124
125}

檢測結果

4.2 攝像頭/視頻中的人臉檢測

face_detector_video.cpp

 1
 2
 3
 4
 5
 6#include <iostream>  
 7#include <cstdlib>  
 8#include <stdio.h>
 9#include <opencv2/opencv.hpp>
10#include <opencv2/dnn.hpp>
11#include <opencv2/dnn/shape_utils.hpp>
12
13using namespace cv;  
14using namespace cv::dnn;  
15using namespace std;  
16const size_t inWidth = 300;  
17const size_t inHeight = 300;  
18const double inScaleFactor = 1.0;  
19const Scalar meanVal(104.0, 177.0, 123.0);  
20
21int main(int argc, char** argv)  
22{  
23    float min_confidence = 0.5;  
24    String modelConfiguration = "face_detector/deploy.prototxt";  
25    String modelBinary = "face_detector/res10_300x300_ssd_iter_140000.caffemodel";  
26    
27    dnn::Net net = readNetFromCaffe(modelConfiguration, modelBinary);  
28    
29    if (net.empty())  
30    {  
31        cerr << "Can't load network by using the following files: " << endl;  
32        cerr << "prototxt:   " << modelConfiguration << endl;  
33        cerr << "caffemodel: " << modelBinary << endl;  
34        cerr << "Models are available here:" << endl;  
35        cerr << "<OPENCV_SRC_DIR>/samples/dnn/face_detector" << endl;  
36        cerr << "or here:" << endl;  
37        cerr << "https://github.com/opencv/opencv/tree/master/samples/dnn/face_detector" << endl;  
38        exit(-1);  
39    }  
40
41    VideoCapture cap(0);  
42    if (!cap.isOpened())  
43    {  
44        cout << "Couldn't open camera : " << endl;  
45        return -1;  
46    }  
47    for (;;)  
48    {  
49        Mat frame;  
50        cap >> frame;
51
52        if (frame.empty())  
53        {  
54            waitKey();  
55            break;  
56        }  
57
58        if (frame.channels() == 4)  
59            cvtColor(frame, frame, COLOR_BGRA2BGR);  
60
61        
62        Mat inputBlob = blobFromImage(frame, inScaleFactor,  
63            Size(inWidth, inHeight), meanVal, false, false);
64                                                            
65
66                                                            
67        net.setInput(inputBlob, "data");
68                                        
69
70                                        
71        Mat detection = net.forward("detection_out");
72                                                      
73
74        vector<double> layersTimings;  
75        double freq = getTickFrequency() / 1000;  
76        double time = net.getPerfProfile(layersTimings) / freq;  
77
78        Mat detectionMat(detection.size[2], detection.size[3], CV_32F, detection.ptr<float>());  
79
80        ostringstream ss;  
81        ss << "FPS: " << 1000 / time << " ; time: " << time << " ms";  
82        putText(frame, ss.str(), Point(20, 20), 0, 0.5, Scalar(0, 0, 255));  
83
84        float confidenceThreshold = min_confidence;  
85        for (int i = 0; i < detectionMat.rows; i++)  
86        {  
87            float confidence = detectionMat.at<float>(i, 2);  
88
89            if (confidence > confidenceThreshold)  
90            {  
91                int xLeftBottom = static_cast<int>(detectionMat.at<float>(i, 3) * frame.cols);  
92                int yLeftBottom = static_cast<int>(detectionMat.at<float>(i, 4) * frame.rows);  
93                int xRightTop = static_cast<int>(detectionMat.at<float>(i, 5) * frame.cols);  
94                int yRightTop = static_cast<int>(detectionMat.at<float>(i, 6) * frame.rows);  
95
96                Rect object((int)xLeftBottom, (int)yLeftBottom,  
97                    (int)(xRightTop - xLeftBottom),  
98                    (int)(yRightTop - yLeftBottom));  
99
100                rectangle(frame, object, Scalar(0, 255, 0));  
101
102                ss.str("");  
103                ss << confidence;  
104                String conf(ss.str());  
105                String label = "Face: " + conf;  
106                int baseLine = 0;  
107                Size labelSize = getTextSize(label, FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);  
108                rectangle(frame, Rect(Point(xLeftBottom, yLeftBottom - labelSize.height),  
109                    Size(labelSize.width, labelSize.height + baseLine)),  
110                    Scalar(255, 255, 255), CV_FILLED);  
111                putText(frame, label, Point(xLeftBottom, yLeftBottom),  
112                    FONT_HERSHEY_SIMPLEX, 0.5, Scalar(0, 0, 0));  
113            }  
114        }  
115        cv::imshow("detections", frame);  
116        if (waitKey(1) >= 0) break;  
117    }  
118    return 0;  
119}  

檢測結果

5 Python版本代碼

最簡單安裝Python版的OpenCV方法

對於OpenCV3.4版本，可直接使用opencv-3.4.1\samples\dnn文件夾中的resnet_ssd_face_python.py；

對於OpenCV3.3.1版本，可參考下述代碼（自己寫的）：

5.1 圖像中的人臉檢測

detect_faces.py

1
2
3
4
5import numpy as np
6import argparse
7import cv2
8
9
10ap = argparse.ArgumentParser()
11ap.add_argument("-i", "--image", required=True,
12    help="path to input image")
13ap.add_argument("-p", "--prototxt", required=True,
14    help="path to Caffe 'deploy' prototxt file")
15ap.add_argument("-m", "--model", required=True,
16    help="path to Caffe pre-trained model")
17ap.add_argument("-c", "--confidence", type=float, default=0.5,
18    help="minimum probability to filter weak detections")
19args = vars(ap.parse_args())
20
21
22print("[INFO] loading model...")
23net = cv2.dnn.readNetFromCaffe(args["prototxt"], args["model"])
24
25
26
27image = cv2.imread(args["image"])
28(h, w) = image.shape[:2]
29blob = cv2.dnn.blobFromImage(cv2.resize(image, (300, 300)), 1.0,
30    (300, 300), (104.0, 177.0, 123.0))
31
32
33
34print("[INFO] computing object detections...")
35net.setInput(blob)
36detections = net.forward()
37
38
39for i in range(0, detections.shape[2]):
40    
41    
42    confidence = detections[0, 0, i, 2]
43
44    
45    
46    if confidence > args["confidence"]:
47        
48        
49        box = detections[0, 0, i, 3:7] * np.array([w, h, w, h])
50        (startX, startY, endX, endY) = box.astype("int")
51
52        
53        
54        text = "{:.2f}%".format(confidence * 100)
55        y = startY - 10 if startY - 10 > 10 else startY + 10
56        cv2.rectangle(image, (startX, startY), (endX, endY),
57            (0, 0, 255), 2)
58        cv2.putText(image, text, (startX, y),
59            cv2.FONT_HERSHEY_SIMPLEX, 0.45, (0, 0, 255), 2)
60
61
62cv2.imshow("Output", image)
63cv2.waitKey(0)

打開cmd命令提示符，切換至路徑下，輸入下述命令：

檢測結果

檢測結果

5.2 攝像頭/視頻中的人臉檢測

detect_faces_video.py

1
2
3
4
5from imutils.video import VideoStream
6import numpy as np
7import argparse
8import imutils
9import time
10import cv2
11
12
13ap = argparse.ArgumentParser()
14ap.add_argument("-p", "--prototxt", required=True,
15    help="path to Caffe 'deploy' prototxt file")
16ap.add_argument("-m", "--model", required=True,
17    help="path to Caffe pre-trained model")
18ap.add_argument("-c", "--confidence", type=float, default=0.5,
19    help="minimum probability to filter weak detections")
20args = vars(ap.parse_args())
21
22
23print("[INFO] loading model...")
24net = cv2.dnn.readNetFromCaffe(args["prototxt"], args["model"])
25
26
27print("[INFO] starting video stream...")
28vs = VideoStream(src=0).start()
29time.sleep(2.0)
30
31
32while True:
33    
34    
35    frame = vs.read()
36    frame = imutils.resize(frame, width=400)
37
38    
39    (h, w) = frame.shape[:2]
40    blob = cv2.dnn.blobFromImage(cv2.resize(frame, (300, 300)), 1.0,
41        (300, 300), (104.0, 177.0, 123.0))
42
43    
44    
45    net.setInput(blob)
46    detections = net.forward()
47
48    
49    for i in range(0, detections.shape[2]):
50        
51        
52        confidence = detections[0, 0, i, 2]
53
54        
55        
56        if confidence < args["confidence"]:
57            continue
58
59        
60        
61        box = detections[0, 0, i, 3:7] * np.array([w, h, w, h])
62        (startX, startY, endX, endY) = box.astype("int")
63
64        
65        
66        text = "{:.2f}%".format(confidence * 100)
67        y = startY - 10 if startY - 10 > 10 else startY + 10
68        cv2.rectangle(frame, (startX, startY), (endX, endY),
69            (0, 0, 255), 2)
70        cv2.putText(frame, text, (startX, y),
71            cv2.FONT_HERSHEY_SIMPLEX, 0.45, (0, 0, 255), 2)
72
73    
74    cv2.imshow("Frame", frame)
75    key = cv2.waitKey(1) & 0xFF
76
77    
78    if key == ord("q"):
79        break
80
81
82cv2.destroyAllWindows()
83vs.stop()

打開cmd命令提示符，切換至路徑下，輸入下述命令：

如果程序出錯，如ImportError: No module named imutils.video。這說明當前Python庫中沒有imutils庫，所以可以使用pip安裝：

檢測結果

總結

本教程介紹並使用了OpenCV最新提供的更加精確的人臉檢測器（與OpenCV的Haar級聯相比）。

這裡的OpenCV人臉檢測器是基於深度學習的，特別是利用ResNet和SSD框架作為基礎網絡。

感謝Aleksandr Rybnikov、OpenCV dnn模塊和Adrian Rosebrock等其他貢獻者的努力，我們可以在自己的應用中享受到這些更加精確的OpenCV人臉檢測器。

為了你的方便，我已經為你準備了本教程所使用的必要文件，請見下述內容。

代碼下載

deep-learning-face-detection.rar：https://anonfile.com/nft4G4d5b1/deep-learning-face-detection.rar

Reference

[1]Face detection with OpenCV and deep learning：https://www.pyimagesearch.com/2018/02/26/face-detection-with-opencv-and-deep-learning/

[2]face_detector：https://github.com/opencv/opencv/tree/master/samples/dnn/face_detector

[3]opencv3.4 發布 dnnFace震撼來襲：http://blog.csdn.net/minstyrain/article/details/78907425

--我是可愛的分割線--      

若喜歡Amusi推送的文章，請掃描下方二維碼關注CVer公眾號！