（OpenCV3.2.0）基于HOG特征的SVM训练与测试实例

本人作为一个机器学习的小菜鸟，在学习相关知识的过程中，CSDN博客给了我很大的帮助，作为回馈，我也来写点博客为后来者铺路。

注意：在程序前需要包含一些必要的OpenCV头文件，我是将这些头文件写在了CvHeadFiles.h文件之中

#include "CvHeadFiles.h"      

#include <iostream>   

#include <fstream>   

#include <string>   

#include <vector> 

using namespace cv;

using namespace cv::ml;

using namespace std;

int main(int argc, char** argv)

{

        //--------------准备训练数据--------------------

        //定义相关变量

        int ImgWidht = 24;

        int ImgHeight = 24;

        vector<string> img_path;

        vector<int> img_catg;

        int nLine = 0;

        string buf;

        int lab;

        ifstream svm_data("SVM_DATA.txt");  

        unsigned long n;

        //依次读入样本的名和标签

        for(;;)

        {

               nLine++;

               if (nLine % 2)

               {

                       svm_data >> buf;

                       if (buf == "end")

                              break;

                       img_path.push_back(buf);//图像路径  

               }

               else

               {

                       svm_data >> lab;

                       img_catg.push_back(lab);

               }

        }

        svm_data.close();//关闭文件   

        //定义相关变量

        Mat data_mat, res_mat;

        int nImgNum = (nLine-1)/2;            //读入样本数量    

        res_mat = Mat::zeros(nImgNum, 1, CV_32SC1);

        Mat src;

        Mat trainImg = Mat::zeros(ImgHeight, ImgWidht, CV_8UC3);  //需要分析的图片   

        //依次提取样本的HOG特征

        for (string::size_type i = 0; i != img_path.size(); i++)

        {

               //读入样本

               src = imread(img_path[i].c_str(), 1);  //依次读入样本图片

               cout << " processing " << img_path[i].c_str() << endl;

               resize(src, trainImg, cv::Size(ImgWidht, ImgHeight), 0, 0, INTER_CUBIC);   //将样本图片调整为标准大小

               

               //对样本提取HOG特征向量

               HOGDescriptor *hog = new HOGDescriptor(cvSize(ImgWidht, ImgHeight), cvSize(16, 16), cvSize(8, 8), cvSize(8, 8), 9);  //具体意思见参考文章1,2      

               vector<float> descriptors;  //存放提取HOG特征向量的结果数组 

               hog->compute(trainImg, descriptors, Size(1, 1), Size(0, 0)); //对读入的样本图片提取HOG特征向量 

               if (i == 0)

               {

                       data_mat = Mat::zeros(nImgNum, descriptors.size(), CV_32FC1); //根据输入图片大小进行分配空间  

               }

               cout << "HOG dims: " << descriptors.size() << endl;

               //存储所提取的样本的HOG特征值

               n = 0;

               for (vector<float>::iterator iter = descriptors.begin();iter != descriptors.end();iter++)

               {

                       data_mat.at<float>(i, n) = *iter;

                       n++;

               }

               

               //存储样本的标签

               res_mat.at<int>(i, 0) = img_catg[i];

               cout << " end processing " << img_path[i].c_str() << " " << img_catg[i] << endl;

        }

        

        //----------------------------设置SVM参数---------------------------------

        cout << "\n------------------------------------------------------------\n";

        cout << "Starting training process" << endl;

        Ptr<SVM> svm = SVM::create();

        svm->setType(SVM::C_SVC);

        svm->setC(10);

        svm->setKernel(SVM::RBF);

        svm->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER, 10000, 1e-6));

        

        //-------------------------------训练SVM------------------------------------------      

        svm->train(data_mat, ROW_SAMPLE, res_mat);   //注意此处data_mat和res_mat的类型

        //利用训练数据和确定的学习参数,进行SVM学习      

        svm->save("SVM_DATA.xml");

        cout << "Finished training process" << endl;

        //--------------------------------检测样本---------------------------------------   

        //依次读入测试样本名

        vector<string> img_tst_path;

        ifstream img_tst("SVM_TEST.txt");

        for(;;)

        {

               img_tst >> buf;

               if (buf == "end")

                       break;

               img_tst_path.push_back(buf);

        }

        img_tst.close();

        //定义相关变量

        Mat test;

        int pos=0, neg=0,sum=0;

        char line[512];

        ofstream predict_txt("SVM_PREDICT.txt");

        for (string::size_type j = 0; j < img_tst_path.size(); j++)

        {

               //依次读入测试样本

               test = imread(img_tst_path[j].c_str(), 1);   //读入图像    

               resize(test, trainImg, cv::Size(ImgWidht, ImgHeight), 0, 0, INTER_CUBIC);   //要搞成同样的大小才可以检测到        

               

               //提取测试样本的HOG特征

               HOGDescriptor *hog = new HOGDescriptor(cvSize(ImgWidht, ImgHeight), cvSize(16, 16), cvSize(8, 8), cvSize(8, 8), 9);  

               vector<float>descriptors;//结果数组      

               hog->compute(trainImg, descriptors, Size(1, 1), Size(0, 0)); //调用计算函数开始计算  

               cout << "The Detection Result:" << endl;

               cout << "HOG dims: " << descriptors.size() << endl;

               //存储所提取的特征

               Mat SVMtrainMat = Mat::zeros(1, descriptors.size(), CV_32FC1);

               n = 0;

               for (vector<float>::iterator iter = descriptors.begin();iter != descriptors.end();iter++)

               {

                       SVMtrainMat.at<float>(0, n) = *iter;

                       n++;

               }

               //对提取的特征进行检验

               int ret = svm->predict(SVMtrainMat);

               if (ret == 1)

                       pos++;

               else

                       neg++;

               std::sprintf(line, "%s %d\r\n", img_tst_path[j].c_str(), ret);

               printf("%s %d\r\n", img_tst_path[j].c_str(), ret);

               //getchar();

               predict_txt << line;

        }

        sum = neg + pos;

        cout << "neg=" << neg << "\n";

        cout << "pos=" << pos << "\n";

        cout << "sum=" << sum << "\n";

        predict_txt.close();

        system("pause");

        return 0;

}