人脸检测库函数的效果比较

年前在OpenCV论坛上看到Shiqi Yu发布的免费的高性能的人脸检测库，于是down下来做了一下效果测试。感兴趣的同学，也可以做一下检测速度的比较。
本人才疏学浅，只是将它与OpenCV自带的人脸检测库函数做了一个粗浅的比较，欢迎各位大神提点。

---

程序功能简介：

程序实现环境：VS2010 + OpenCV 2.4.11
程序退出：在窗口中敲击回车键即可；
旋转图像：鼠标拖动窗口中的Angle轴滑块即可；
缩放图像：鼠标拖动窗口中的Scale轴滑块即可；
平移图像：在窗口中任意位置，鼠标左键拖拽图像即可；
平移图像后复原：在窗口中任意位置，单击鼠标右键即可；
切换人脸检测库函数：鼠标拖动窗口中的Mode轴滑块即可。

对于Mode值的说明：

• 当$0\le Mode\le 4$时，检测函数采用的是OpenCV自带的级联检测模块CascadeClassifier（如何使用级联分类器做人脸检测，可以参见OpenCV的Tutorial，这里有相应的中文翻译版）。具体说明如下：

当mode=0时，加载的是lbpcascade_frontalface.xml；
当mode=1时，加载的是haarcascade_frontalface_alt.xml；
当mode=2时，加载的是haarcascade_frontalface_alt2.xml；
当mode=3时，加载的是haarcascade_frontalface_default.xml；
当mode=4时，加载的是haarcascade_frontalface_alt_tree.xml；

• 当$5\le Mode\le 8$时，检测函数采用的是Shiqi Yu发布的[免费的高性能的人脸检测库]（使用方法可参见它自带的示例程序）。具体说明如下：

当mode=5时，调用的是facedetect_frontal()；
当mode=6时，调用的是facedetect_frontal_surveillance()；
当mode=7时，调用的是facedetect_multiview()；
当mode=8时，调用的是facedetect_multiview_reinforce()；

---

程序代码：

#include <time.h>
#include <stdio.h>
#include <vector>
#include <iostream>
#include <opencv.hpp>
#include "facedetect-dll.h"

using namespace cv;
using namespace std;

#pragma comment(lib,"libfacedetect.lib")
//#pragma comment(lib,"libfacedetect-x64.lib")

//Do not change the buffer size!
#define DETECT_BUFFER_SIZE 0xC004

static Mat onRot(int pos, int rat, Mat* im);
static Mat onAffine(Mat* im, Point2f dif);
static void onMouse(int Event, int x, int y, int drag, Mat* im);
static int Detect_Display(Mat& im, int mod);

void main(int argc, char* argv[])
{
    cout<<"Hello, VS10. I'm back!"<<endl;
    Mat Pic = imread("Friends2.jpg"), Cam(Pic), dst;

    namedWindow("Video");   int pos = 36, rat = 10, mod = 0;
    createTrackbar("Angle","Video", &pos,72);
    createTrackbar("Scale","Video", &rat,40);
    createTrackbar("Mode","Video", &mod,8);
    setMouseCallback("Video", (MouseCallback)onMouse, &Cam);

    //VideoCapture cap(0);
    while((waitKey(100)&255)!=13) //Enter=13
    {
        dst = onRot(pos, rat, &Cam);
        Detect_Display(dst, mod);
        imshow("Video", dst);
    }//end while
}//end main

static int Detect_Display(Mat& im, int mod)
{
    Mat gray;   cvtColor(im, gray, CV_BGR2GRAY); //must gray
    vector<Rect> faces, eyes;   equalizeHist(gray, gray); //normalization

    int *pResults = NULL; //!!! DO NOT RELEASE pResults !!!
    //pBuffer is used in the facedetect_XXX functions:
    unsigned char* pBuffer = (unsigned char*)malloc(DETECT_BUFFER_SIZE);
    if(!pBuffer)    {fprintf(stderr, "Can not alloc buffer.\n");    return 0;}
    CascadeClassifier Eyes("haarcascade_eye_tree_eyeglasses.xml"), Face;

    switch(mod)
    {
    case 0: if(!Face.load("lbpcascade_frontalface.xml"))    return 0;
        Face.detectMultiScale(gray, faces, 1.1, 2, CV_HAAR_SCALE_IMAGE|0, Size(30,30)); break;
    case 1: if(!Face.load("haarcascade_frontalface_alt.xml"))   return 0;
        Face.detectMultiScale(gray, faces, 1.1, 2, CV_HAAR_SCALE_IMAGE|0, Size(30,30)); break;
    case 2: if(!Face.load("haarcascade_frontalface_alt2.xml"))  return 0;
        Face.detectMultiScale(gray, faces, 1.1, 2, CV_HAAR_SCALE_IMAGE|0, Size(30,30)); break;
    case 3: if(!Face.load("haarcascade_frontalface_default.xml"))   return 0;
        Face.detectMultiScale(gray, faces, 1.1, 2, CV_HAAR_SCALE_IMAGE|0, Size(30,30)); break;
    case 4: if(!Face.load("haarcascade_frontalface_alt_tree.xml"))  return 0;
        Face.detectMultiScale(gray, faces, 1.1, 2, CV_HAAR_SCALE_IMAGE|0, Size(30,30)); break;
    //frontal face detection: it's fast, but cannot detect side view faces.
    case 5: pResults = facedetect_frontal(pBuffer, (unsigned char*)(gray.ptr(0)),
                                  gray.cols, gray.rows, (int)gray.step, 1.2f, 2, 48);   break;
    //frontal face detection for video surveillance: it can detect faces with bad illumination.
    case 6: pResults = facedetect_frontal_surveillance(pBuffer, (unsigned char*)(gray.ptr(0)),
                                  gray.cols, gray.rows, (int)gray.step, 1.2f, 2, 48);   break;
    //multiview face detection: it can detect side view faces, but slower than facedetect_frontal().
    case 7: pResults = facedetect_multiview(pBuffer, (unsigned char*)(gray.ptr(0)),
                                  gray.cols, gray.rows, (int)gray.step, 1.2f, 2, 48);   break;
    //reinforced multiview face detection: it's better but slower than facedetect_multiview().
    case 8: pResults = facedetect_multiview_reinforce(pBuffer, (unsigned char*)(gray.ptr(0)),
                                  gray.cols, gray.rows, (int)gray.step, 1.2f, 2, 48);   break;
    }//end switch
    if(mod>4)   faces.resize(pResults?(*pResults):0); //for facedetect_XXX functions

    cout<<faces.size()<<" faces detected!\n";
    for(int i=0; i<faces.size(); ++i) //mark each face
    {
        if(mod>4) //for facedetect_XXX functions
        {   short* p = ((short*)(pResults+1)) + 6*i;
            faces[i].x = p[0];  faces[i].width = p[2];
            faces[i].y = p[1];  faces[i].height = p[3];
            cout<<"face_rect = "<<faces[i]<<", neighbors = "<<p[4]<<".\n";
        }//end if
        rectangle(im, faces[i], Scalar(255,0,255), 2); //mark with rectangle

        Mat faceROI = gray(faces[i]); //In each face, detect eyes
        Eyes.detectMultiScale(faceROI, eyes, 1.1, 2, CV_HAAR_SCALE_IMAGE|0, Size(30,30));
        for(int j=0; j<eyes.size(); ++j) //mark each eye
        {
            Point center(faces[i].x + eyes[j].x + eyes[j].width*0.5, faces[i].y + eyes[j].y + eyes[j].height*0.5);
            int radius = cvRound((eyes[j].width + eyes[j].height)/4);
            circle(im, center, radius, Scalar(255,0,0), 2); //mark with circle
        }//end for
    }//end for
    free(pBuffer);  return faces.size(); //release pBuffer
}//end Detect_Display

static Mat onRot(int pos, int rat, Mat* im)
{
    Point center(im->cols/2, im->rows/2);
    double angle = 5*pos-180, scale = 0.1*rat;
    Mat rot = getRotationMatrix2D(center, angle, scale), dst;
    warpAffine(*im, dst, rot, im->size());  return dst;
}//end onRot

static void onMouse(int Event, int x, int y, int drag, Mat* im)
{ //only effective for Picture, not fit for Cam.
    static Mat res = im->clone();   static const Point ini(0,0);
    static Point seed(ini), tdf(ini);   Point now(x,y);

    switch(Event)
    {
    case CV_EVENT_LBUTTONDOWN:  seed = now; break;
    case CV_EVENT_LBUTTONUP:    tdf += (now-seed);  break;
    case CV_EVENT_RBUTTONUP:    res.copyTo(*im);    tdf = ini;  break;
    }//end switch

    if(drag==CV_EVENT_FLAG_LBUTTON) *im = onAffine(&res, tdf+(now-seed));
}//end onMouse

static Mat onAffine(Mat* im, Point2f dif)
{
    Point2f org[3],tri[3];  org[1] = Point2f(im->cols/2, 0);
    org[0] = Point2f(0,0);  org[2] = Point2f(0, im->rows/2);
    for(int i=0; i<3; ++i)  tri[i] = org[i] + dif;

    Mat warp = getAffineTransform(org, tri), dst;
    warpAffine(*im, dst, warp, im->size()); return dst;
}//end onAffine

---

程序效果图：

---

---

---

---

---

---

---

---