实践|OpenCV4.2使用DNN进行人脸检测一（图片篇）

res10_300x300_ssd_iter_140000_fp16.caffemodel
deploy.prototxt

https://raw.githubusercontent.com/opencv/opencv_3rdparty/dnn_samples_face_detector_20180205_fp16/res10_300x300_ssd_iter_140000_fp16.caffemodel

opencv_face_detector_uint8.pb
opencv_face_detector.pbtxt

https://raw.githubusercontent.com/opencv/opencv_3rdparty/dnn_samples_face_detector_20180220_uint8/opencv_face_detector_uint8.pb

//Tensorflow
cv::dnn::readNetFromTensorflow(_modelbinary, _modeldesc)

//Caffe
cv::dnn::readNetFromCaffe(_modelbinary, _modeldesc)

Mat cv::dnn::blobFromImage(
    InputArray     image,  //输入图像
    double     scalefactor = 1.0,   //图像缩放的比率
    const Size &     size = Size(),  //返回的Mat中数据的尺寸
    const Scalar &     mean = Scalar(),    //关于mean参数，如果之前没有深入研究过深度学习，这个还是不太好理解的。首先给出mean的数值：(104 117 123);数字从什么地方来的呢？这个是在Net训练的时候设定的，可以看到在训练的时候transform_param中设置了mean
    bool     swapRB = false,  //是否交换R和B分量
    bool     crop = false,  //裁剪标志，指示是否在调整大小后裁剪图像
    int     ddepth = CV_32F //图像的数据类型，目前仅支持32F和8U
    )

void cv::dnn::Net::setInput    (    
    InputArray     blob,  //上 个函数blobFromImage的返回值
    const String &     name = "",  //输入图层的名称
    double     scalefactor = 1.0,  //可选的标准化比例
    const Scalar &     mean = Scalar() //可选的平均减法值
    )

 Mat forward(const String& outputName = String())
 
 //outputName：需要输出的图层的名称
 //返回：指定图层outputName的第一个输出的blob。默认情况下，为整个网络运行正向传递。注意：返回Mat类型，这是一个4D数，rows and cols can only hold 2 dimensions, so they are not used here, and set to -1

#pragma once

#include<opencv2/opencv.hpp>
#include<opencv2/dnn/dnn.hpp>

using namespace std;
using namespace cv;

class dnnfacedetect
{
private:
  string _modelbinary, _modeldesc;
  dnn::Net _net;
public:
  //构造函数 传入模型文件
  dnnfacedetect();
  dnnfacedetect(string modelBinary, string modelDesc);

  ~dnnfacedetect();
  //置信阈值
  float confidenceThreshold;
  double inScaleFactor;
  size_t inWidth;
  size_t inHeight;
  Scalar meanVal;

  //初始化DNN网络
  bool initdnnNet();

  //人脸检测
  vector<Mat> detect(Mat frame);
};

#include "dnnfacedetect.h"


dnnfacedetect::dnnfacedetect()
{
  dnnfacedetect("", "");
}

//构造函数
dnnfacedetect::dnnfacedetect(string modelBinary, string modelDesc)
{
  _modelbinary = modelBinary;
  _modeldesc = modelDesc;

  //初始化置信阈值
  confidenceThreshold = 0.6;
  inScaleFactor = 0.5;
  inWidth = 300;
  inHeight = 300;
  meanVal = Scalar(104.0, 177.0, 123.0);
}

dnnfacedetect::~dnnfacedetect()
{
  _net.~Net();
}

//初始化dnnnet
bool dnnfacedetect::initdnnNet()
{
  _net = dnn::readNetFromTensorflow(_modelbinary, _modeldesc);
  _net.setPreferableBackend(dnn::DNN_BACKEND_OPENCV);
  _net.setPreferableTarget(dnn::DNN_TARGET_CPU);

  return !_net.empty();
}

//人脸检测
vector<Mat> dnnfacedetect::detect(Mat frame)
{
  Mat tmpsrc = frame;
  vector<Mat> dsts = vector<Mat>();
  // 修改通道数
  if (tmpsrc.channels() == 4)
    cvtColor(tmpsrc, tmpsrc, COLOR_BGRA2BGR);
  // 输入数据调整
  Mat inputBlob = dnn::blobFromImage(tmpsrc, inScaleFactor,
    Size(inWidth, inHeight), meanVal, false, false);
  _net.setInput(inputBlob, "data");

  //人脸检测
  Mat detection = _net.forward("detection_out");

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

  //检测出的结果进行绘制和存放到dsts中
  for (int i = 0; i < detectionMat.rows; i++) {
    //置值度获取
    float confidence = detectionMat.at<float>(i, 2);
    //如果大于阈值说明检测到人脸
    if (confidence > confidenceThreshold) {
      //计算矩形
      int xLeftBottom = static_cast<int>(detectionMat.at<float>(i, 3) * tmpsrc.cols);
      int yLeftBottom = static_cast<int>(detectionMat.at<float>(i, 4) * tmpsrc.rows);
      int xRightTop = static_cast<int>(detectionMat.at<float>(i, 5) * tmpsrc.cols);
      int yRightTop = static_cast<int>(detectionMat.at<float>(i, 6) * tmpsrc.rows);
      //生成矩形
      Rect rect((int)xLeftBottom, (int)yLeftBottom,
        (int)(xRightTop - xLeftBottom),
        (int)(yRightTop - yLeftBottom));

      //截出图矩形存放到dsts数组中
      Mat tmp = tmpsrc(rect);
      dsts.push_back(tmp);

      //在原图上用红框画出矩形
      rectangle(frame, rect, Scalar(0, 0, 255));
    }
  }

  return dsts;
}

#include<opencv2/opencv.hpp>
#include<iostream>
#include <direct.h>
#include "dnnfacedetect.h"

using namespace std;
using namespace cv;

int main(int argc, char* argv) {
  //获取程序目录
  char filepath[256];
  _getcwd(filepath, sizeof(filepath));

  cout << filepath << endl;
  //定义模型文件
  string ModelBinary = (string)filepath + "/opencv_face_detector_uint8.pb";
  string ModelDesc = (string)filepath + "/opencv_face_detector.pbtxt";

  //图片文件
  string picdesc = (string)filepath + "/lena.jpg";

  cout << ModelBinary << endl;
  cout << ModelDesc << endl;

  //加载图片
  Mat frame = imread(picdesc);
  imshow("src", frame);

  try
  {
    //初始化
    dnnfacedetect fdetect = dnnfacedetect(ModelBinary, ModelDesc);
    if (!fdetect.initdnnNet())
    {
      cout << "初始化DNN人脸检测失败！" << endl;
      return -1;
    }

    if (!frame.empty()) {
      vector<Mat> dst = fdetect.detect(frame);
      if (!dst.empty()) {
        for (int i = 0; i < dst.size(); i++) {
          string title = "dst" + i;
          imshow(title, dst[i]);
        }
        imshow("src2", frame);
      }
    }
  }
  catch (const std::exception & ex)
  {
    cout << ex.what() << endl;
  }

  waitKey(0);
  return 0;
}