前言
项目有一个模块需要将不同类别的图片进行分类,共有三个类别,使用SVM实现分类。
实现步骤:
1.创建训练样本库;
2.训练、测试SVM模型;
3.SVM的数据要求;
实现系统:
windows_x64、opencv2.4.10、 VS2013
实现过程:
1.创建训练样本库;
1)将图片以包含类别的名称进行命名,比如0(1).jpg等等;
2)将所有已命名正确的训练样本保存在同一个文件夹中;
3)在训练样本库的文件夹目录下创建python源文件;
python代码:
import sys import os import string import re if __name__=='__main__': print('Begin generate path and label.') path_file=open('train_path.txt','w') path='E:/carriage_recognition/redplate_detection/svm_train_test/data/train/model' pic_type='.png' pat=re.compile(r'^(\d+)') files=os.listdir(path) files_tmp=[] for i in files: if pic_type in i and not os.path.isdir(path+'/'+i): files_tmp.append(i) files=files_tmp for file in range(len(files)): ret=pat.match(files[file]) path_file.write(path+'/'+files[file]+'\n') if file: path_file.write(ret.group(1)+'\n') else: path_file.write(ret.group(1)) path_file.close() print('finish......')
4)运行代码,即可,生成包含图片名称和类别的文本文件,用于SVM训练过程中读物图片获取相应的类别标签;
E:/carriage_recognition/redplate_detection/svm_train_test/data/train/model/0 (1).png 0 E:/carriage_recognition/redplate_detection/svm_train_test/data/train/model/0 (10).png 0
奇数行表示训练样本图片的路径名称;偶数行表示该图片的类别标签;
2.训练、测试SVM模型;
1)image.h,主要实现过程的代码;
#include
#include
#include
#include //红牌事件检测头文件
#include
using namespace std;
using namespace cv;
#define ON_STUDY 0
#define Num 3 //类别数目
#define STANDARD_ROW 65
#define STANDARD_COL 85
#define STANDARD_ROW_CHOOSE 65
#define STANDARD_COL_CHOOSE 85
#define CHANELS 1
class NumTrainData
{
public:
NumTrainData()
{
memset(data, 0, sizeof(data));
result = -1;
}
public:
float data[CHANELS*STANDARD_COL_CHOOSE*STANDARD_ROW_CHOOSE];
int result;
};
vector img_path;//输入文件名变量
vector img_test_path;//输入文件名变量
vector img_catg;
vector img_test_catg;
int nLine = 0;
string buf;
unsigned long n;
vector buffer;
int featureLen = CHANELS*STANDARD_COL_CHOOSE*STANDARD_ROW_CHOOSE;
char* test_path = "./test_path.txt";
char* train_path = "./train_path.txt";
//存放输出结果
char* save_path = "./SVM_DATA_train_0.5_0.2.xml";
ofstream matrix_config("./fusion_matrix_0.5_0.2.txt"); //存放混淆矩阵
string save_wrong_results = "./wrong_0.5_0.2"; //存放识别错误的结果
void ReadTrainData()
{
ifstream svm_data(train_path);//训练样本图片的路径都写在这个txt文件中,使用python可以得到这个txt文件
while (svm_data)//将训练样本文件依次读取进来
{
if (getline(svm_data, buf))
{
nLine++;
if (nLine % 2 == 0)//注:奇数行是图片全路径,偶数行是标签
{
img_catg.push_back(atoi(buf.c_str()));//atoi将字符串转换成整型,标志(0,1,2,...,9),注意这里至少要有两个类别,否则会出错
}
else
{
img_path.push_back(buf);//图像路径
}
}
}
svm_data.close();//关闭文件
}
void ReadTestData()
{
ifstream svm_data(test_path);//训练样本图片的路径都写在这个txt文件中,使用python可以得到这个txt文件
while (svm_data)//将训练样本文件依次读取进来
{
if (getline(svm_data, buf))
{
nLine++;
if (nLine % 2 == 0)//注:奇数行是图片全路径,偶数行是标签
{
img_test_catg.push_back(atoi(buf.c_str()));//atoi将字符串转换成整型,标志(0,1,2,...,9),注意这里至少要有两个类别,否则会出错
}
else
{
img_test_path.push_back(buf);//图像路径
}
}
}
svm_data.close();//关闭文件
}
void LoadTrainData()
{
Mat src; //= Mat::zeros(rows, cols, CV_8UC1);
Mat dst;
NumTrainData rtd;
cout << "Begin load training data...." << endl;
for (int i = 0; i < img_path.size(); i++)
{
rtd.result = img_catg[i];
int k = 0;
if (CHANELS == 1) // gray image
{
src = imread(img_path[i].c_str(), 0);
dst = src;
Mat temp = Mat::zeros(STANDARD_ROW, STANDARD_COL, CV_8UC1);
//尺寸归一化
resize(dst, temp, temp.size());
float m[CHANELS*STANDARD_COL_CHOOSE*STANDARD_ROW_CHOOSE];
for (int i = 0; i(i, j);
}
}
}
else if (CHANELS == 3) // 3-channel image
{
src = imread(img_path[i].c_str(), 1);
dst = src;
Mat temp = Mat::zeros(STANDARD_ROW, STANDARD_COL, CV_8UC1); //大小归一化
resize(dst, temp, temp.size());
//cout << temp.channels() << endl;
for (int i = 0; i < STANDARD_ROW_CHOOSE; i++)
{
for (int j = 0; j < STANDARD_COL_CHOOSE; j++)
{
Vec3b& mp = temp.at(i, j);
float B = mp.val[0];
//cout << "B=" << B << endl;
float G = mp.val[1];
//cout << "G=" << B << endl;
float R = mp.val[2];
//cout << "R=" << B << endl;
rtd.data[k++] = B; //R
rtd.data[k++] = G; //G
rtd.data[k++] = R; //B
}
}
}
buffer.push_back(rtd);
//cout << i << "th Image is loaded!" << endl;
}
cout << "Loading image finished!" << endl;
}
void SVMPredict()
{
int x = 0;
//_mkdir(save_test_preprocess.c_str());
_mkdir(save_wrong_results.c_str());
int fusion_matrix[Num][Num] = { 0 };
CvSVM svm;
svm.load(save_path);
Mat src,dst;
Mat m = Mat::zeros(1, featureLen, CV_32FC1);
NumTrainData rtd;
int label = -1;
int right = 0, error = 0;
save_wrong_results += "/%d_true_%d_false_%d.png";
//
double ptrue_rtrue = 0;
double ptrue = 0;
double rtrue = 0;
//
for (int i = 0; i < img_test_path.size(); i++)
{
label = img_test_catg[i];
rtd.result = label;
if (CHANELS == 1)
{
src = imread(img_test_path[i].c_str(), 0);
dst = src;
Mat temp = Mat::zeros(STANDARD_ROW, STANDARD_COL, CV_8UC1); //大小归一化
resize(dst, temp, temp.size());
for (int i = 0; i(0, j + i * STANDARD_COL) = temp.at(i, j);
}
}
normalize(m, m);
}
else if (CHANELS == 3) // 3-channel image
{
src = imread(img_test_path[i].c_str(), 1);
dst = src;
Mat temp = Mat::zeros(STANDARD_ROW, STANDARD_COL, CV_8UC1); //大小归一化
resize(dst, temp, temp.size());
int k = 0;
for (int i = 0; i < STANDARD_ROW_CHOOSE; i++)
{
for (int j = 0; j < STANDARD_COL_CHOOSE; j++)
{
Vec3b& mp = temp.at(i, j);
float B = mp.val[0];
float G = mp.val[1];
float R = mp.val[2];
m.at(0, k++) = B; //R
m.at(0, k++) = G; //G
m.at(0, k++) = R; //B
}
}
}
int ret = svm.predict(m);
//if (ret == 3)
// ret = 1;
cout << "Picture->" << img_test_path[i].c_str() << " : \nTrue label is [" << label << "] Predicted label is [" << ret << "]" << endl;
//
//计算FSCORE指标各个参数
if (label == 0 && ret == 0) ptrue_rtrue++;//识别为红牌且实际为红牌;
if (ret == 0) ptrue++;//识别为红牌的个数
if (label == 0) rtrue++;//实际为红牌的个数
//
//存储错误图片
if (label != ret)
{
x++;
char filename[200];
src = imread(img_test_path[i].c_str(), 1);
sprintf(filename, save_wrong_results.c_str(), x, label, ret);
imwrite(filename, src);
}
//计算混淆矩阵
//fusion_matrix[label][ret] = fusion_matrix[label][ret] + 1;
}
//
//FSCORE
std::cout << "count_all: " << img_test_path.size() << std::endl;
std::cout << "ptrue_rtrue: " << ptrue_rtrue << std::endl;
std::cout << "ptrue: " << ptrue << std::endl;
std::cout << "rtrue: " << rtrue << std::endl;
//precise
double precise = 0;
if (ptrue != 0)
{
precise = ptrue_rtrue / ptrue;
std::cout << "precise: " << precise << std::endl;
}
else
{
std::cout << "precise: " << "NA" << std::endl;
}
//recall
double recall = 0;
if (rtrue != 0)
{
recall = ptrue_rtrue / rtrue;
std::cout << "recall: " << recall << std::endl;
}
else
{
std::cout << "recall: " << "NA" << std::endl;
}
//FSCORE
double FScore = 0;
if (precise + recall != 0)
{
FScore = 2 * (precise * recall) / (precise + recall);
std::cout << "FScore: " << FScore << std::endl;
}
else
{
std::cout << "FScore: " << "NA" << std::endl;
}
//
//for (size_t i = 0; i < Num; i++)
//{
// for (size_t j = 0; j < Num; j++)
// {
// matrix_config << fusion_matrix[i][j] << " ";
// }
// matrix_config << endl;
//}
//matrix_config.close();
cout << "Task finished!output_matix" << endl;
getchar();
}
void SVMTrain(vector& trainData)
{
int testCount = trainData.size();
Mat m = Mat::zeros(1, featureLen, CV_32FC1);
Mat data = Mat::zeros(testCount, featureLen, CV_32FC1);
//Mat res = Mat::zeros(testCount, 1, CV_32SC1);
Mat res = Mat::zeros(testCount, 1, CV_32SC1);
for (int i = 0; i< testCount; i++)
{
NumTrainData td = trainData.at(i);
memcpy(m.data, td.data, featureLen * sizeof(float));
normalize(m, m);
memcpy(data.data + i*featureLen * sizeof(float), m.data, featureLen * sizeof(float));
cout << td.result << endl;
res.at(i, 0) = td.result;
}
/////////////START SVM TRAINNING//////////////////
//CvSVM svm = CvSVM();
CvSVM svm;
CvSVMParams param;
CvTermCriteria criteria;
criteria = cvTermCriteria(CV_TERMCRIT_EPS, 1000, FLT_EPSILON);
param = CvSVMParams(CvSVM::C_SVC, CvSVM::RBF, 10.0, 0.5, 1.0, 0.2, 0.5, 0.1, NULL, criteria); //gamma=2;C=3
cout << "Begin to train model using given train data.....\n Total training sample count is " << testCount << endl;
svm.train(data, res, Mat(), Mat(), param);
svm.save(save_path);
cout << "Finish" << endl;
}
2)主要函数说明;
2.1)SVMTrain函数主要实现模型的训练,其中训练参数使用RBF核,主要调整gamma和C这两个参数,固定一个参数调整另一个参数,最后确定模型参数分别为0.5/0.2;
2.2)SVMPredict函数主要实现对测试样本库的测试,并使用FScore指标测试SVM模型的性能;也可以使用混淆矩阵测试性能;
2.3)ReadTrainData/ ReadTestData函数分别用于获取训练和测试样本库图片的名称和类别标签;
2.4)LoadTrainData函数用于读取训练数据,并进行图像处理;
2.5)代码中使用整张图片的信息进行归一化之后作为特征;
3)主函数入口
#include "image.h" int main(int argc, char *argv[]) { #if (ON_STUDY) ReadTrainData(); LoadTrainData(); SVMTrain(buffer); #else ReadTestData(); SVMPredict(); #endif getchar(); }
参数ON_STUDY表示选择进行训练或者测试的标志位;
3.SVM的数据要求;
需要说明的是就是SVM对于输入的数据类型是有要求的,即mTrainData(训练数据矩阵)以及mFlagPosNeg(标签矩阵)都必须为CV_32FC1类型(我的环境标签矩阵是CV_32SC1类型的),因此需要进行类型转换,而且必须保证转换完之后数值都不能大于1,这就给我们了两点启示:1)不能直接用下采样后的图像像素作为训练数据的输入,需要进行类型的归一化。2)类型转换时要使用normlize函数,保证其数值范围不大于1,而不能简单的使用Mat的成员函数coverto,只变类型不变数值范围。( 需要注意!)
问题:
该实现过程需要人工调整参数,比较繁琐,可以思考一下,是否还存在其他问题;
参考:
1.http://blog.csdn.net/firefight/article/details/6452188
2.opencv中SVM的那些事儿
完