C++中使用OPENCV对深度学习的特征图进行可视化

佚名 6年前 (2019-05-05) 随笔 511人围观抢沙发百度已收录

#include <iostream>
#include <unistd.h>
#include <opencv4/opencv2/opencv.hpp>//理论上opencv3.3.*和3.4.*也可以，未测试
#include <opencv4/opencv2/dnn.hpp>
#include <opencv4/opencv2/dnn/dnn.hpp>
using namespace std;

std::string strip(const std::string& str,const std::string& del){
    std::string res = str;
    size_t pos=res.find_first_of(del);
    size_t len=del.length();
    while(pos!=std::string::npos){
        res.erase(res.begin()+pos,res.begin()+pos+len);
        pos=res.find_first_of(del);
    }
    return res;
}

int main(int argc, char *argv[])
{
    //建议 OPENCV 4.*.*
    cerr<<"WORKING IN OPENCV_VERSION "<<CV_VERSION<<'\n';
    cv::Mat src = cv::imread("src.jpg");

    //加载模型所用的网络结构文件和参数文件，这里用的是caffe训练得到的MobileNet-SSD
    string cfg     ="MobileNetSSD_deploy.prototxt";
    string weights ="mobilenet_iter_25000.caffemodel";
    cv::dnn::Net net = cv::dnn::readNetFromCaffe(cfg,weights);

    if(net.empty()){
        cerr<<"loaded net failed.\n";
        return -1;
    }

    //图片输入到输入层
    cv::Mat inputBlob = cv::dnn::blobFromImage(src, 1/255.F, cv::Size(300, 300), cv::Scalar(), false, false);
    net.setInput(inputBlob,"data");

    //遍历一遍网络结构，查看各层的变化（暂未输出各层的连接关系）
    vector<cv::String> outputname=net.getLayerNames();
    cv::dnn::MatShape netInputSize = {1,3,300,300};
    vector<cv::dnn::MatShape> netlastSize;
    vector<vector<cv::dnn::MatShape> >layerSizes;
    for(size_t i=0;i<outputname.size();++i){
        vector<cv::dnn::MatShape>inputLayerSize;
        vector<cv::dnn::MatShape>outputLayerSize;
        net.getLayerShapes(netInputSize,i,inputLayerSize,outputLayerSize);
        cerr<<"layer <<"<<outputname[i]<<">> size [\n";
        for(size_t j=0;j<inputLayerSize.size();++j){
            cerr<<"\t["<<inputLayerSize[j][0];for(size_t k=1;k<inputLayerSize[j].size();++k)cerr<<" x "<<inputLayerSize[j][k];cerr<<"]\n";
        }
        cerr<<"\tto\n";
        for(size_t j=0;j<outputLayerSize.size();++j){
            cerr<<"\t["<<outputLayerSize[j][0];for(size_t k=1;k<outputLayerSize[j].size();++k)cerr<<" x "<<outputLayerSize[j][k];cerr<<"]\n";
        }
        cerr<<"]\n";
        layerSizes.push_back(outputLayerSize);
        netlastSize=inputLayerSize;
    }

    //进行预测
    vector<vector<cv::Mat> > outputBlob;
    try{
        cerr<<"\nNet forward ";
        net.forward(outputBlob, outputname);
        cerr<<"done\n\n";
    }catch(exception e){
        cerr << e.what() << '\n';
    }

    //要保存的层的名字，具体名字可以在之前输出的信息里面找
    string selectLayer="conv4";

    for(size_t i=1;i<outputname.size();++i){
        if(selectLayer!=outputname[i])continue;
        for(size_t j=0;j<outputBlob[i].size();++j){
            try{

                cv::Mat blob = outputBlob[i-1][j].clone();
                cerr<<"blob dim = "<<blob.size.dims()<<" with type "<<blob.type()<<'\n';
                //dim 4: b * c * w * h
                //dim 3: b * w * h
                //dim 2: b * w

                cv::Mat saveimg, temp;
                float* data=(float*)blob.data;
                string savepath;
                int s[2]={0};

                //保存不同dim情况下的特征图
                switch(blob.size.dims()){
                case 2://单轴
                    saveimg.create(layerSizes[i][j][0],1,CV_8UC1);
                    for(size_t pi=0; pi<layerSizes[i][j][0]; ++pi){
                        saveimg.data[pi]=data[pi];
                    }
                    cv::normalize(saveimg,saveimg,255,0,cv::NORM_MINMAX);
                    savepath = "layers/layer_"+strip(outputname[i],"/")+".jpg";
                    cv::imwrite(savepath,saveimg);
                    saveimg.release();
                    break;
                case 3://单层卷积特征图
                    s[1]=layerSizes[i][j][1];
                    s[0]=layerSizes[i][j][2];
                    saveimg.create(layerSizes[i][j][1],layerSizes[i][j][2],CV_8UC1);
                    for(size_t pi=0; pi<s[1]*s[0]; ++pi){
                        saveimg.data[pi]=(uchar)data[pi];
                    }
                    cv::normalize(saveimg,saveimg,255,0,cv::NORM_MINMAX);
                    savepath = "layers/layer_"+strip(outputname[i],"/")+".jpg";
                    cv::imwrite(savepath,saveimg);
                    saveimg.release();
                    break;
                case 4://多层卷积特征图
                    for(int k=0;k<layerSizes[i][j][1];++k){
                        s[1]=layerSizes[i][j][2];
                        s[0]=layerSizes[i][j][3];
                        saveimg.create(layerSizes[i][j][2],layerSizes[i][j][3],CV_8UC1);
                        for(size_t pi=0; pi<s[1]*s[0]; ++pi){
                            saveimg.data[pi]=(uchar)data[k*s[1]*s[0]+pi];
                        }
                        cv::normalize(saveimg,saveimg,255,0,cv::NORM_MINMAX);
                        savepath = "layers/layer_"+strip(outputname[i],"/")+"_channal_"+to_string(k)+".jpg";
                        cv::imwrite(savepath,saveimg);
                        saveimg.release();
                    }
                    break;
                default:
                    cerr<<" exordinary dim\n";
                    break;
                }
                cerr<<"save to "<<savepath<<'\n';

                blob.release();
            }catch(exception e){
                cerr<<" escape beacuse "<<e.what()<<'\n';
            }
        }
    }

    //检验数据类型的尺寸和序号
    fprintf(stderr, "float %d, double %d\n",sizeof(float),sizeof(double));
    //blob <CV_32FC1,5> float
    fprintf(stderr, "type list : <CV_8UC1,%d> <CV_8UC3,%d> <CV_16FC1,%d> <CV_16FC3,%d> <CV_16SC1,%d> <CV_16SC3,%d> <CV_32FC1,%d> <CV_32SC1,%d> ",
            CV_8UC1,CV_8UC3,CV_16FC1,CV_16FC3,CV_16SC1,CV_16SC3,CV_32FC1,CV_32SC1);
    return 0;
}