在本教程中，我们将讨论各种人脸检测方法，并对各种方法进行比较。下面是主要的人脸检测方法：

1 OpenCV中的Haar Cascade人脸分类器；

2 OpenCV中的深度学习人脸分类器；

3 Dlib中的hog人脸分类器；

4 Dlib中的深度学习人脸分类器。

Dlib是一个C++工具包（也有python版本），代码地址： http://dlib.net/

本文不涉及任何原理，只讲具体的应用。所有代码模型见：

https://download.csdn.net/download/luohenyj/10997489

https://github.com/luohenyueji/OpenCV-Practical-Exercise

如果没有积分（系统自动设定资源分数）看看参考链接。我搬运过来的，大修改没有。pch是预编译文件。Opencv版本3.4.3以上。

1 OpenCV中的Haar Cascade人脸分类器

基于Haar Cascade的人脸检测器自2001年提出以来，一直是人脸检测领域的研究热点。这种模型和其变种在这里找到：

https://github.com/opencv/opencv/tree/master/data/haarcascades

这种方法优点在CPU上几乎是实时工作的，方法简单可以在不同的尺度上检测人脸。实际就是一个级联分类器，参数可以调整，网上有相关资料。但是不管怎么调整误报率很高，而且人脸框选结果不是那么准确。

代码

C++:

#include "pch.h"

#include "face_detection.h"

/**

 * @brief 人脸检测haar级联

 *

 * @param frame 原图

 * @param faceCascadePath 模型文件

 * @return Mat

 */

Mat detectFaceHaar(Mat frame, string faceCascadePath)

{

	//图像缩放

	auto inHeight = 300;

	auto inWidth = 0;

	if (!inWidth)

	{

		inWidth = (int)(((float)frame.cols / (float)frame.rows) * inHeight);

	}

	resize(frame, frame, Size(inWidth, inHeight));

	//转换为灰度图

	Mat frameGray = frame.clone();

	//cvtColor(frame, frameGray, CV_BGR2GRAY);

	//级联分类器

	CascadeClassifier faceCascade;

	faceCascade.load(faceCascadePath);

	std::vector<Rect> faces;

	faceCascade.detectMultiScale(frameGray, faces);

	for (size_t i = 0; i < faces.size(); i++)

	{

		int x1 = faces[i].x;

		int y1 = faces[i].y;

		int x2 = faces[i].x + faces[i].width;

		int y2 = faces[i].y + faces[i].height;

		Rect face_rect(Point2i(x1, y1), Point2i(x2, y2));

		rectangle(frameGray, face_rect, cv::Scalar(0, 255, 0), 2, 4);

	}

	return frameGray;

}

python:

from __future__ import division

import cv2

import time

import sys

def detectFaceOpenCVHaar(faceCascade, frame, inHeight=300, inWidth=0):

    frameOpenCVHaar = frame.copy()

    frameHeight = frameOpenCVHaar.shape[0]

    frameWidth = frameOpenCVHaar.shape[1]

    if not inWidth:

        inWidth = int((frameWidth / frameHeight) * inHeight)

    scaleHeight = frameHeight / inHeight

    scaleWidth = frameWidth / inWidth

    frameOpenCVHaarSmall = cv2.resize(frameOpenCVHaar, (inWidth, inHeight))

    frameGray = cv2.cvtColor(frameOpenCVHaarSmall, cv2.COLOR_BGR2GRAY)

    faces = faceCascade.detectMultiScale(frameGray)

    bboxes = []

    for (x, y, w, h) in faces:

        x1 = x

        y1 = y

        x2 = x + w

        y2 = y + h

        cvRect = [int(x1 * scaleWidth), int(y1 * scaleHeight),

                  int(x2 * scaleWidth), int(y2 * scaleHeight)]

        bboxes.append(cvRect)

        cv2.rectangle(frameOpenCVHaar, (cvRect[0], cvRect[1]), (cvRect[2], cvRect[3]), (0, 255, 0),

                      int(round(frameHeight / 150)), 4)

    return frameOpenCVHaar, bboxes

if __name__ == "__main__" :

    source = 0

    if len(sys.argv) > 1:

        source = sys.argv[1]

    faceCascade = cv2.CascadeClassifier('./haarcascade_frontalface_default.xml')

    cap = cv2.VideoCapture(source)

    hasFrame, frame = cap.read()

    vid_writer = cv2.VideoWriter('output-haar-{}.avi'.format(str(source).split(".")[0]),cv2.VideoWriter_fourcc('M','J','P','G'), 15, (frame.shape[1],frame.shape[0]))

    frame_count = 0

    tt_opencvHaar = 0

    while(1):

        hasFrame, frame = cap.read()

        if not hasFrame:

            break

        frame_count += 1

        t = time.time()

        outOpencvHaar, bboxes = detectFaceOpenCVHaar(faceCascade, frame)

        tt_opencvHaar += time.time() - t

        fpsOpencvHaar = frame_count / tt_opencvHaar

        label = "OpenCV Haar ; FPS : {:.2f}".format(fpsOpencvHaar)

        cv2.putText(outOpencvHaar, label, (10, 50), cv2.FONT_HERSHEY_SIMPLEX, 1.4, (0, 0, 255), 3, cv2.LINE_AA)

        cv2.imshow("Face Detection Comparison", outOpencvHaar)

        vid_writer.write(outOpencvHaar)

        if frame_count == 1:

            tt_opencvHaar = 0

        k = cv2.waitKey(10)

        if k == 27:

            break

    cv2.destroyAllWindows()

    vid_writer.release()

2 OpenCV中的深度学习人脸分类器

OpenCV3.3以上版本就有该分类器的模型。模型来自论文： https://arxiv.org/abs/1512.02325

但是提供了两种不同的模型。一种是16位浮点数的caffe人脸模型(5.4MB)，另外一种是8bit量化后的tensorflow人脸模型(2.7MB)。量化是指比如可以用0~255表示原来32个bit所表示的精度，通过牺牲精度来降低每一个权值所需要占用的空间。通常情况深度学习模型会有冗余计算量，冗余性决定了参数个数。因此合理的量化网络也可保证精度的情况下减小模型的存储体积，不会对网络的精度造成影响。具体可以看看深度学习fine-

tuning的论文。通常这种操作可以稍微降低精度，提高速度，大大减少模型体积。

这种方法速度慢了点，但是精度不错。对于调用模型代码写的很清楚。但是tensorflow模型有点小问题，可能只能在opencv3.4.3以上版本通过readNet函数调用。

代码

C++:

#include "pch.h"

#include "face_detection.h"

//检测图像宽高

const size_t inWidth = 300;

const size_t inHeight = 300;

//缩放比例

const double inScaleFactor = 1.0;

//阈值

const double confidenceThreshold = 0.7;

//均值

const cv::Scalar meanVal(104.0, 177.0, 123.0);

/**

 * @brief 人脸检测Opencv ssd

 *

 * @param frame 原图

 * @param configFile 模型结构定义文件

 * @param weightFile 模型文件

 * @return Mat

 */

Mat detectFaceOpenCVDNN(Mat frame, string configFile, string weightFile)

{

	Mat frameOpenCVDNN = frame.clone();

	Net net;

	Mat inputBlob;

	int frameHeight = frameOpenCVDNN.rows;

	int frameWidth = frameOpenCVDNN.cols;

	//获取文件后缀

	string suffixStr = configFile.substr(configFile.find_last_of('.') + 1);

	//判断是caffe模型还是tensorflow模型

	if (suffixStr == "prototxt")

	{

		net = dnn::readNetFromCaffe(configFile, weightFile);

		inputBlob = cv::dnn::blobFromImage(frameOpenCVDNN, inScaleFactor, cv::Size(inWidth, inHeight), meanVal, false, false);

	}

	else

	{

		//bug

		//net = dnn::readNetFromTensorflow(configFile, weightFile);

		net = dnn::readNet(configFile, weightFile);

		inputBlob = cv::dnn::blobFromImage(frameOpenCVDNN, inScaleFactor, cv::Size(inWidth, inHeight), meanVal, true, false);

	}

	//读图检测

	net.setInput(inputBlob, "data");

	cv::Mat detection = net.forward("detection_out");

	cv::Mat detectionMat(detection.size[2], detection.size[3], CV_32F, detection.ptr<float>());

	for (int i = 0; i < detectionMat.rows; i++)

	{

		//分类精度

		float confidence = detectionMat.at<float>(i, 2);

		if (confidence > confidenceThreshold)

		{

			//左上角坐标

			int x1 = static_cast<int>(detectionMat.at<float>(i, 3) * frameWidth);

			int y1 = static_cast<int>(detectionMat.at<float>(i, 4) * frameHeight);

			//右下角坐标

			int x2 = static_cast<int>(detectionMat.at<float>(i, 5) * frameWidth);

			int y2 = static_cast<int>(detectionMat.at<float>(i, 6) * frameHeight);

			//画框

			cv::rectangle(frameOpenCVDNN, cv::Point(x1, y1), cv::Point(x2, y2), cv::Scalar(0, 255, 0), 2, 4);

		}

	}

	return frameOpenCVDNN;

}

python

from __future__ import division

import cv2

import time

import sys

def detectFaceOpenCVDnn(net, frame):

    frameOpencvDnn = frame.copy()

    frameHeight = frameOpencvDnn.shape[0]

    frameWidth = frameOpencvDnn.shape[1]

    blob = cv2.dnn.blobFromImage(frameOpencvDnn, 1.0, (300, 300), [104, 117, 123], False, False)

    net.setInput(blob)

    detections = net.forward()

    bboxes = []

    for i in range(detections.shape[2]):

        confidence = detections[0, 0, i, 2]

        if confidence > conf_threshold:

            x1 = int(detections[0, 0, i, 3] * frameWidth)

            y1 = int(detections[0, 0, i, 4] * frameHeight)

            x2 = int(detections[0, 0, i, 5] * frameWidth)

            y2 = int(detections[0, 0, i, 6] * frameHeight)

            bboxes.append([x1, y1, x2, y2])

            cv2.rectangle(frameOpencvDnn, (x1, y1), (x2, y2), (0, 255, 0), int(round(frameHeight/150)), 8)

    return frameOpencvDnn, bboxes

if __name__ == "__main__" :

    # OpenCV DNN supports 2 networks.

    # 1. FP16 version of the original caffe implementation ( 5.4 MB )

    # 2. 8 bit Quantized version using Tensorflow ( 2.7 MB )

    DNN = "TF"

    if DNN == "CAFFE":

        modelFile = "models/res10_300x300_ssd_iter_140000_fp16.caffemodel"

        configFile = "models/deploy.prototxt"

        net = cv2.dnn.readNetFromCaffe(configFile, modelFile)

    else:

        modelFile = "models/opencv_face_detector_uint8.pb"

        configFile = "models/opencv_face_detector.pbtxt"

        net = cv2.dnn.readNetFromTensorflow(modelFile, configFile)

    conf_threshold = 0.7

    source = 0

    if len(sys.argv) > 1:

        source = sys.argv[1]

    cap = cv2.VideoCapture(source)

    hasFrame, frame = cap.read()

    vid_writer = cv2.VideoWriter('output-dnn-{}.avi'.format(str(source).split(".")[0]),cv2.VideoWriter_fourcc('M','J','P','G'), 15, (frame.shape[1],frame.shape[0]))

    frame_count = 0

    tt_opencvDnn = 0

    while(1):

        hasFrame, frame = cap.read()

        if not hasFrame:

            break

        frame_count += 1

        t = time.time()

        outOpencvDnn, bboxes = detectFaceOpenCVDnn(net,frame)

        tt_opencvDnn += time.time() - t

        fpsOpencvDnn = frame_count / tt_opencvDnn

        label = "OpenCV DNN ; FPS : {:.2f}".format(fpsOpencvDnn)

        cv2.putText(outOpencvDnn, label, (10,50), cv2.FONT_HERSHEY_SIMPLEX, 1.4, (0, 0, 255), 3, cv2.LINE_AA)

        cv2.imshow("Face Detection Comparison", outOpencvDnn)

        vid_writer.write(outOpencvDnn)

        if frame_count == 1:

            tt_opencvDnn = 0

        k = cv2.waitKey(10)

        if k == 27:

            break

    cv2.destroyAllWindows()

    vid_writer.release()