MTCNN人脸及特征点检测--基于树莓派3B+及ncnn架构

概述

本文尝试在树莓派3B+上用ncnn框架测试MTCNN。

ncnn的基本编译和使用请参考《在树莓派3B+上编译ncnn并用benchmark和mobilenet_yolo测试》。本文在这个博客基础上进行操作。

操作步骤

下载mtcnn

从mtcn-ncnn项目中下载mtcnn子目录，把这个目录放在最新的ncnn源码目录下

添加对mtcnn的支持

修改ncnn最顶层的CMakeList.txt,增加对mtcnn的支持

add_subdirectory(examples)
add_subdirectory(benchmark)
add_subdirectory(src)
# 添加对mtcnn目录的支持
add_subdirectory(mtcnn)

重命名mtcnn.cpp

将mtcnn/mtcnn.cpp删除，将mtcnn_new.cpp重命名为mtcnn.cpp

修改mtcnn.cpp源码,把imgproc.hpp那行修改下：

// #include <opencv2/imgproc.hpp>
#include <opencv2/imgproc/imgproc.hpp>

修改ncnn-root-dir/mtcnn/CMakeList.txt, 去掉imgcodecs：

# find_package(OpenCV REQUIRED core highgui imgproc imgcodecs)
find_package(OpenCV REQUIRED core highgui imgproc)

如果不这么改，编译报错：opencv_imgcodecs is required but was not found

因为我本地的opencv库是2.4的，所以不需要imgcodecs这个选项。

编译

$ cd <ncnn-root-dir>
$ sudo mkdir -p build
$ cd build
$ cmake -DCMAKE_TOOLCHAIN_FILE=../toolchains/pi3.toolchain.cmake -DPI3=ON ..
$ make -j4 # 生成./src/libncnn.a

生成的二进制文件为ncnn-root-dir/build/mtcnn/mtcnn, 将它复制到ncnn-root-dir/mtcnn目录下。

测试mtcnn

mtcn-ncnn这个项目意见转换好了mtcnn的ncnn模型放在在ncnn-root-dir/mtcnn，所以这里就直接调用了，我没有单独做模型转换。

$ cd <ncnn-root-dir>/mtcnn
# ./mtcnn 1.jpg

测试结果存入result.jpg，取下来看就可以了。

时间测试

测试过程发现，第一次测试的时间会长一些，第二次及以后的时间会缩短。

我在上篇文章在树莓派3B+上编译ncnn并用benchmark和mobilenet_yolo测试介绍过默认开启openmp加速的方案。本次编译mtcnn也是这么操作的。也就是说我的测试同时使用了neon和openmp加速。

简单汇总（时间是估的，没精确算）：

分辨率	1920x1080（8个人脸）	648x610（2个人脸）	324x305（2个人脸）
平均耗时（ms）	80	14	3.3

mtcnn.cpp修改后的源码

从上面可以知道，最终编译出的二进制文件是由原始项目里的mtcnn_new.cpp修改的，但是原始文件只测试一次，我需要做多次测试求平均，所以我对这个文件进行了简单修改，每次测试100次，目视就可以估算平均了。

#include <stdio.h>
#include <algorithm>
#include <vector>
#include <math.h>
#include <iostream>
#include <sys/time.h>
#include <opencv2/core/core.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>

#include "net.h"
using namespace std;
using namespace cv;

struct Bbox
{
    float score;
    int x1;
    int y1;
    int x2;
    int y2;
    float area;
    bool exist;
    float ppoint[10];
    float regreCoord[4];
};

struct orderScore
{
    float score;
    int oriOrder;
};
bool cmpScore(orderScore lsh, orderScore rsh){
    if(lsh.score<rsh.score)
        return true;
    else
        return false;
}
static float getElapse(struct timeval *tv1,struct timeval *tv2)
{
    float t = 0.0f;
    if (tv1->tv_sec == tv2->tv_sec)
        t = (tv2->tv_usec - tv1->tv_usec)/1000.0f;
    else
        t = ((tv2->tv_sec - tv1->tv_sec) * 1000 * 1000 + tv2->tv_usec - tv1->tv_usec)/1000.0f;
    return t;
}

class mtcnn{
public:
    mtcnn();
    void detect(ncnn::Mat& img_, std::vector<Bbox>& finalBbox);
private:
    void generateBbox(ncnn::Mat score, ncnn::Mat location, vector<Bbox>& boundingBox_, vector<orderScore>& bboxScore_, float scale);
    void nms(vector<Bbox> &boundingBox_, std::vector<orderScore> &bboxScore_, const float overlap_threshold, string modelname="Union");
    void refineAndSquareBbox(vector<Bbox> &vecBbox, const int &height, const int &width);

    ncnn::Net Pnet, Rnet, Onet;
    ncnn::Mat img;

    const float nms_threshold[3] = {0.5, 0.7, 0.7};
    const float threshold[3] = {0.6, 0.6, 0.6};
    const float mean_vals[3] = {127.5, 127.5, 127.5};
    const float norm_vals[3] = {0.0078125, 0.0078125, 0.0078125};
    std::vector<Bbox> firstBbox_, secondBbox_,thirdBbox_;
    std::vector<orderScore> firstOrderScore_, secondBboxScore_, thirdBboxScore_;
    int img_w, img_h;
};

mtcnn::mtcnn(){
    Pnet.load_param("det1.param");
    Pnet.load_model("det1.bin");
    Rnet.load_param("det2.param");
    Rnet.load_model("det2.bin");
    Onet.load_param("det3.param");
    Onet.load_model("det3.bin");
}

void mtcnn::generateBbox(ncnn::Mat score, ncnn::Mat location, std::vector<Bbox>& boundingBox_, std::vector<orderScore>& bboxScore_, float scale){
    int stride = 2;
    int cellsize = 12;
    int count = 0;
    //score p
    float *p = score.channel(1);
    float *plocal = location.channel(0);
    Bbox bbox;
    orderScore order;
    for(int row=0;row<score.h;row++){
        for(int col=0;col<score.w;col++){
            if(*p>threshold[0]){
                bbox.score = *p;
                order.score = *p;
                order.oriOrder = count;
                bbox.x1 = round((stride*col+1)/scale);
                bbox.y1 = round((stride*row+1)/scale);
                bbox.x2 = round((stride*col+1+cellsize)/scale);
                bbox.y2 = round((stride*row+1+cellsize)/scale);
                bbox.exist = true;
                bbox.area = (bbox.x2 - bbox.x1)*(bbox.y2 - bbox.y1);
                for(int channel=0;channel<4;channel++)
                    bbox.regreCoord[channel]=location.channel(channel)[0];
                boundingBox_.push_back(bbox);
                bboxScore_.push_back(order);
                count++;
            }
            p++;
            plocal++;
        }
    }
}
void mtcnn::nms(std::vector<Bbox> &boundingBox_, std::vector<orderScore> &bboxScore_, const float overlap_threshold, string modelname){
    if(boundingBox_.empty()){
        return;
    }
    std::vector<int> heros;
    //sort the score
    sort(bboxScore_.begin(), bboxScore_.end(), cmpScore);

    int order = 0;
    float IOU = 0;
    float maxX = 0;
    float maxY = 0;
    float minX = 0;
    float minY = 0;
    while(bboxScore_.size()>0){
        order = bboxScore_.back().oriOrder;
        bboxScore_.pop_back();
        if(order<0)continue;
        if(boundingBox_.at(order).exist == false) continue;
        heros.push_back(order);
        boundingBox_.at(order).exist = false;//delete it

        for(int num=0;num<boundingBox_.size();num++){
            if(boundingBox_.at(num).exist){
                //the iou
                maxX = (boundingBox_.at(num).x1>boundingBox_.at(order).x1)?boundingBox_.at(num).x1:boundingBox_.at(order).x1;
                maxY = (boundingBox_.at(num).y1>boundingBox_.at(order).y1)?boundingBox_.at(num).y1:boundingBox_.at(order).y1;
                minX = (boundingBox_.at(num).x2<boundingBox_.at(order).x2)?boundingBox_.at(num).x2:boundingBox_.at(order).x2;
                minY = (boundingBox_.at(num).y2<boundingBox_.at(order).y2)?boundingBox_.at(num).y2:boundingBox_.at(order).y2;
                //maxX1 and maxY1 reuse 
                maxX = ((minX-maxX+1)>0)?(minX-maxX+1):0;
                maxY = ((minY-maxY+1)>0)?(minY-maxY+1):0;
                //IOU reuse for the area of two bbox
                IOU = maxX * maxY;
                if(!modelname.compare("Union"))
                    IOU = IOU/(boundingBox_.at(num).area + boundingBox_.at(order).area - IOU);
                else if(!modelname.compare("Min")){
                    IOU = IOU/((boundingBox_.at(num).area<boundingBox_.at(order).area)?boundingBox_.at(num).area:boundingBox_.at(order).area);
                }
                if(IOU>overlap_threshold){
                    boundingBox_.at(num).exist=false;
                    for(vector<orderScore>::iterator it=bboxScore_.begin(); it!=bboxScore_.end();it++){
                        if((*it).oriOrder == num) {
                            (*it).oriOrder = -1;
                            break;
                        }
                    }
                }
            }
        }
    }
    for(int i=0;i<heros.size();i++)
        boundingBox_.at(heros.at(i)).exist = true;
}
void mtcnn::refineAndSquareBbox(vector<Bbox> &vecBbox, const int &height, const int &width){
    if(vecBbox.empty()){
        cout<<"Bbox is empty!!"<<endl;
        return;
    }
    float bbw=0, bbh=0, maxSide=0;
    float h = 0, w = 0;
    float x1=0, y1=0, x2=0, y2=0;
    for(vector<Bbox>::iterator it=vecBbox.begin(); it!=vecBbox.end();it++){
        if((*it).exist){
            bbw = (*it).x2 - (*it).x1 + 1;
            bbh = (*it).y2 - (*it).y1 + 1;
            x1 = (*it).x1 + (*it).regreCoord[0]*bbw;
            y1 = (*it).y1 + (*it).regreCoord[1]*bbh;
            x2 = (*it).x2 + (*it).regreCoord[2]*bbw;
            y2 = (*it).y2 + (*it).regreCoord[3]*bbh;

            w = x2 - x1 + 1;
            h = y2 - y1 + 1;
          
            maxSide = (h>w)?h:w;
            x1 = x1 + w*0.5 - maxSide*0.5;
            y1 = y1 + h*0.5 - maxSide*0.5;
            (*it).x2 = round(x1 + maxSide - 1);
            (*it).y2 = round(y1 + maxSide - 1);
            (*it).x1 = round(x1);
            (*it).y1 = round(y1);

            //boundary check
            if((*it).x1<0)(*it).x1=0;
            if((*it).y1<0)(*it).y1=0;
            if((*it).x2>width)(*it).x2 = width - 1;
            if((*it).y2>height)(*it).y2 = height - 1;

            it->area = (it->x2 - it->x1)*(it->y2 - it->y1);
        }
    }
}
void mtcnn::detect(ncnn::Mat& img_, std::vector<Bbox>& finalBbox_){
    firstBbox_.clear();
    firstOrderScore_.clear();
    secondBbox_.clear();
    secondBboxScore_.clear();
    thirdBbox_.clear();
    thirdBboxScore_.clear();
    
    img = img_;
    img_w = img.w;
    img_h = img.h;
    img.substract_mean_normalize(mean_vals, norm_vals);

    float minl = img_w<img_h?img_w:img_h;
    int MIN_DET_SIZE = 12;
    int minsize = 90;
    float m = (float)MIN_DET_SIZE/minsize;
    minl *= m;
    float factor = 0.709;
    int factor_count = 0;
    vector<float> scales_;
    while(minl>MIN_DET_SIZE){
        if(factor_count>0)m = m*factor;
        scales_.push_back(m);
        minl *= factor;
        factor_count++;
    }
    orderScore order;
    int count = 0;

    for (size_t i = 0; i < scales_.size(); i++) {
        int hs = (int)ceil(img_h*scales_[i]);
        int ws = (int)ceil(img_w*scales_[i]);
        //ncnn::Mat in = ncnn::Mat::from_pixels_resize(image_data, ncnn::Mat::PIXEL_RGB2BGR, img_w, img_h, ws, hs);
        ncnn::Mat in;
        resize_bilinear(img_, in, ws, hs);
        //in.substract_mean_normalize(mean_vals, norm_vals);
        ncnn::Extractor ex = Pnet.create_extractor();
        ex.set_light_mode(true);
        //ex.set_num_threads(4);
        ex.input("data", in);
        ncnn::Mat score_, location_;
        ex.extract("prob1", score_);
        ex.extract("conv4-2", location_);
        std::vector<Bbox> boundingBox_;
        std::vector<orderScore> bboxScore_;
        generateBbox(score_, location_, boundingBox_, bboxScore_, scales_[i]);
        nms(boundingBox_, bboxScore_, nms_threshold[0]);

        for(vector<Bbox>::iterator it=boundingBox_.begin(); it!=boundingBox_.end();it++){
            if((*it).exist){
                firstBbox_.push_back(*it);
                order.score = (*it).score;
                order.oriOrder = count;
                firstOrderScore_.push_back(order);
                count++;
            }
        }
        bboxScore_.clear();
        boundingBox_.clear();
    }
    //the first stage's nms
    if(count<1)return;
    nms(firstBbox_, firstOrderScore_, nms_threshold[0]);
    refineAndSquareBbox(firstBbox_, img_h, img_w);
    printf("firstBbox_.size()=%d\n", firstBbox_.size());

    //second stage
    count = 0;
    for(vector<Bbox>::iterator it=firstBbox_.begin(); it!=firstBbox_.end();it++){
        if((*it).exist){
            ncnn::Mat tempIm;
            copy_cut_border(img, tempIm, (*it).y1, img_h-(*it).y2, (*it).x1, img_w-(*it).x2);
            ncnn::Mat in;
            resize_bilinear(tempIm, in, 24, 24);
            ncnn::Extractor ex = Rnet.create_extractor();
            ex.set_light_mode(true);
            //ex.set_num_threads(4);
            ex.input("data", in);
            ncnn::Mat score, bbox;
            ex.extract("prob1", score);
            ex.extract("conv5-2", bbox);
            if((score[1])>threshold[1]){
                for(int channel=0;channel<4;channel++)
                    it->regreCoord[channel]=bbox[channel];
                it->area = (it->x2 - it->x1)*(it->y2 - it->y1);
                it->score = score[1];
                secondBbox_.push_back(*it);
                order.score = it->score;
                order.oriOrder = count++;
                secondBboxScore_.push_back(order);
            }
            else{
                (*it).exist=false;
            }
        }
    }
    printf("secondBbox_.size()=%d\n", secondBbox_.size());
    if(count<1)return;
    nms(secondBbox_, secondBboxScore_, nms_threshold[1]);
    refineAndSquareBbox(secondBbox_, img_h, img_w);

    //third stage 
    count = 0;
    for(vector<Bbox>::iterator it=secondBbox_.begin(); it!=secondBbox_.end();it++){
        if((*it).exist){
            ncnn::Mat tempIm;
            copy_cut_border(img, tempIm, (*it).y1, img_h-(*it).y2, (*it).x1, img_w-(*it).x2);
            ncnn::Mat in;
            resize_bilinear(tempIm, in, 48, 48);
            ncnn::Extractor ex = Onet.create_extractor();
            ex.set_light_mode(true);
            //ex.set_num_threads(4);
            ex.input("data", in);
            ncnn::Mat score, bbox, keyPoint;
            ex.extract("prob1", score);
            ex.extract("conv6-2", bbox);
            ex.extract("conv6-3", keyPoint);
            if(score[1]>threshold[2]){
                for(int channel=0;channel<4;channel++)
                    it->regreCoord[channel]=bbox[channel];
                it->area = (it->x2 - it->x1)*(it->y2 - it->y1);
                it->score = score[1];
                for(int num=0;num<5;num++){
                    (it->ppoint)[num] = it->x1 + (it->x2 - it->x1)*keyPoint[num];
                    (it->ppoint)[num+5] = it->y1 + (it->y2 - it->y1)*keyPoint[num+5];
                }

                thirdBbox_.push_back(*it);
                order.score = it->score;
                order.oriOrder = count++;
                thirdBboxScore_.push_back(order);
            }
            else
                (*it).exist=false;
            }
        }

    printf("thirdBbox_.size()=%d\n", thirdBbox_.size());
    if(count<1)return;
    refineAndSquareBbox(thirdBbox_, img_h, img_w);
    nms(thirdBbox_, thirdBboxScore_, nms_threshold[2], "Min");
    finalBbox_ = thirdBbox_;
}

int main(int argc, char** argv)
{
    const char* imagepath = argv[1];

    cv::Mat cv_img = cv::imread(imagepath, CV_LOAD_IMAGE_COLOR);
    if (cv_img.empty())
    {
        fprintf(stderr, "cv::imread %s failed\n", imagepath);
        return -1;
    }
    std::vector<Bbox> finalBbox;
    mtcnn mm;
    ncnn::Mat ncnn_img = ncnn::Mat::from_pixels(cv_img.data, ncnn::Mat::PIXEL_BGR2RGB, cv_img.cols, cv_img.rows);
    struct timeval  tv1,tv2;
    struct timezone tz1,tz2;
    
    # 测试100次
    for (int cnt=0; cnt<100; cnt++) {
        gettimeofday(&tv1,&tz1);
        mm.detect(ncnn_img, finalBbox);
        gettimeofday(&tv2,&tz2);
        printf( "%s = %g ms \n ", "Detection All time", getElapse(&tv1, &tv2) );
    }
    for(vector<Bbox>::iterator it=finalBbox.begin(); it!=finalBbox.end();it++){
        if((*it).exist){
            rectangle(cv_img, Point((*it).x1, (*it).y1), Point((*it).x2, (*it).y2), Scalar(0,0,255), 2,8,0);
            for(int num=0;num<5;num++)circle(cv_img,Point((int)*(it->ppoint+num), (int)*(it->ppoint+num+5)),3,Scalar(0,255,255), -1);
        }
    }
    imwrite("result.jpg",cv_img);
    return 0;
}

参考资料

在树莓派3B+上编译ncnn并用benchmark和mobilenet_yolo测试

MTCNN人脸及特征点检测—代码应用详解（基于ncnn架构）

mtcn-ncnn