环境:win10+opencv3.4.4+C++
1、准备单字符数据,数字:
这里是生成数据,生成参考:https://blog.51cto.com/u_8681773/6004679
2、使用脚本生成tran.txt和test.txt列表(注意文件编码格式utf-8):
import os
import cv2
import glob
import pathlib
import random
from os import listdir, getcwd
from os.path import join
# 将程序生成的数字数据整理到train.txt文件中,train.txt中的格式如下:D:\dataset\icdar2017rctw\recognition\train\image_0_0.jpg 金氏眼镜 Latin
# 图片路径
data_path = r'D:\datasets\gen_number_str_char_boxes\test'
# 写入到文件
trainfile = r'D:\datasets\gen_number_str_char_boxes\test.txt'
if __name__ == '__main__':
file_list = glob.glob(data_path + '/*.jpg', recursive=True)
train_file = open(trainfile, 'a', encoding='utf-8-sig') # 带BOM的UTF-8格式
for file_obj in file_list:
file_path = os.path.join(data_path, file_obj)
file_name, file_extend = os.path.splitext(file_obj)
d = pathlib.Path(file_obj)
imagename = str(d.stem)
lineData = imagename.split('_')
image_label = lineData[1]
train_file.write(file_path + ' ' + image_label + '\n')
train_file.close()
3、开始训练:
#include<iostream>
#include<opencv.hpp>
#include <string>
#include <fstream>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <opencv2/objdetect/objdetect.hpp>
#include <opencv2/ml/ml.hpp>
//参考:https://www.cnblogs.com/xuanyuyt/p/6405944.html
using namespace std;
using namespace cv;
int main()
{
vector<string> img_path;//输入文件名变量
vector<int> img_catg;
int nLine = 0;
string line;
size_t pos;
ifstream svm_data("./train.txt");//训练样本图片的路径都写在这个txt文件中,使用bat批处理文件可以得到这个txt文件
unsigned long n;
while (svm_data)//将训练样本文件依次读取进来
{
if (getline(svm_data, line))
{
nLine++;
pos = line.find_last_of(' ');
img_path.push_back(line.substr(0, pos));//图像路径
img_catg.push_back(atoi(line.substr(pos + 1).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
}
}
svm_data.close();//关闭文件
int nImgNum = nLine; //nImgNum是样本数量,只有文本行数的一半,另一半是标签
cv::Mat data_mat(nImgNum, 324, CV_32FC1);//第二个参数,即矩阵的列是由下面的descriptors的大小决定的,可以由descriptors.size()得到,且对于不同大小的输入训练图片,这个值是不同的
data_mat.setTo(cv::Scalar(0));
//类型矩阵,存储每个样本的类型标志
cv::Mat res_mat(nImgNum, 1, CV_32S);
res_mat.setTo(cv::Scalar(0));
cv::Mat src;
cv::Mat trainImg(cv::Size(28, 28), 8, 3);//需要分析的图片,这里默认设定图片是28*28大小,所以上面定义了324,如果要更改图片大小,可以先用debug查看一下descriptors是多少,然后设定好再运行
//处理HOG特征
for (string::size_type i = 0; i != img_path.size(); i++)
{
src = cv::imread(img_path[i].c_str(), 1);
if (src.data == NULL)//if (src == NULL)
{
cout << " can not load the image: " << img_path[i].c_str() << endl;
continue;
}
//cout << " 处理: " << img_path[i].c_str() << endl;
cv::resize(src, trainImg, trainImg.size());
cv::HOGDescriptor* hog = new cv::HOGDescriptor(cv::Size(28, 28), cv::Size(14, 14), cv::Size(7, 7), cv::Size(7, 7), 9);
vector<float>descriptors;//存放结果
hog->compute(trainImg, descriptors, cv::Size(1, 1), cv::Size(0, 0)); //Hog特征计算
//cout << "HOG dims: " << descriptors.size() << endl;
n = 0;
for (vector<float>::iterator iter = descriptors.begin(); iter != descriptors.end(); iter++)
{
//cvmSet(data_mat, i, n, *iter);
data_mat.at<float>(i, n) = *iter;//存储HOG特征
n++;
}
//cvmSet(res_mat, i, 0, img_catg[i]);
res_mat.at<int>(i, 0) = img_catg[i];
//cout << " 处理完毕: " << img_path[i].c_str() << " " << img_catg[i] << endl;
}
cout << "computed features!" << endl;
cv::Ptr<cv::ml::SVM> svm = cv::ml::SVM::create();//新建一个SVM
svm->setType(cv::ml::SVM::C_SVC);
svm->setKernel(cv::ml::SVM::LINEAR);
svm->setC(1);
//-------------------不使用参数优化-------------------------//
svm->setTermCriteria(cv::TermCriteria(CV_TERMCRIT_EPS, 1000, FLT_EPSILON));
svm->train(data_mat, cv::ml::ROW_SAMPLE, res_mat);//训练数据
//-------------------参数优化-------------------------//
//svm->setTermCriteria = cv::TermCriteria(cv::TermCriteria::MAX_ITER, (int)1e7, 1e-6);
//cv::Ptr<cv::ml::TrainData> td = cv::ml::TrainData::create(data_mat, cv::ml::ROW_SAMPLE, res_mat);
//svm->trainAuto(td, 10);
//保存训练好的分类器
svm->save("HOG_SVM_DATA.xml");
cout << "saved model!" << endl;
//检测样本
cv::Mat test;//IplImage *test;
char result[512];
vector<string> img_test_path;
vector<int> img_test_catg;
int coorect = 0;
ifstream img_tst("./test.txt"); //加载需要预测的图片集合,这个文本里存放的是图片全路径,不要标签
while (img_tst)
{
if (getline(img_tst, line))
{
pos = line.find_last_of(' ');
img_test_catg.push_back(atoi(line.substr(pos + 1).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
img_test_path.push_back(line.substr(0, pos));//图像路径
}
}
img_tst.close();
ofstream predict_txt("SVM_PREDICT.txt");//把预测结果存储在这个文本中
for (string::size_type j = 0; j != img_test_path.size(); j++)//依次遍历所有的待检测图片
{
test = cv::imread(img_test_path[j].c_str(), 1);
if (test.data == NULL)//test == NULL
{
cout << " can not load the image: " << img_test_path[j].c_str() << endl;
continue;
}
cv::Mat trainTempImg(cv::Size(28, 28), 8, 3);
trainTempImg.setTo(cv::Scalar(0));
cv::resize(test, trainTempImg, trainTempImg.size());
cv::HOGDescriptor* hog = new cv::HOGDescriptor(cv::Size(28, 28), cv::Size(14, 14), cv::Size(7, 7), cv::Size(7, 7), 9);
vector<float>descriptors;//结果数组
hog->compute(trainTempImg, descriptors, cv::Size(1, 1), cv::Size(0, 0));
//cout << "HOG dims: " << descriptors.size() << endl;
cv::Mat SVMtrainMat(1, descriptors.size(), CV_32FC1);
int n = 0;
for (vector<float>::iterator iter = descriptors.begin(); iter != descriptors.end(); iter++)
{
SVMtrainMat.at<float>(0, n) = *iter;
n++;
}
int ret = svm->predict(SVMtrainMat);//检测结果
if (ret == img_test_catg[j])
coorect++;
sprintf(result, "%s %d\r\n", img_test_path[j].c_str(), ret);
predict_txt << result; //输出检测结果到文本
}
predict_txt.close();
cout << coorect * 100.0 / img_test_path.size() << "%" << endl;
return 0;
}
4、开始测试:
#include<iostream>
#include<opencv.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <opencv2/objdetect/objdetect.hpp>
#include <opencv2/ml/ml.hpp>
using namespace std;
using namespace cv;
using namespace cv::ml;
int main(int argc, char* argv[])
{
cv::Ptr<cv::ml::SVM> svm = cv::ml::SVM::create();
svm = cv::ml::SVM::load("HOG_SVM_DATA.xml");;//加载训练好的xml文件,这里训练的是10K个手写数字
//检测样本
cv::Mat test;
char result[300]; //存放预测结果
test = cv::imread("./img/100293_3_2895.jpg", 1); //待预测图片,用系统自带的画图工具随便手写
if (!test.data)
{
return -1;
}
cv::Mat trainTempImg(cv::Size(28, 28), 8, 3);
trainTempImg.setTo(cv::Scalar(0));
cv::resize(test, trainTempImg, trainTempImg.size());
cv::HOGDescriptor* hog = new cv::HOGDescriptor(cv::Size(28, 28), cv::Size(14, 14), cv::Size(7, 7), cv::Size(7, 7), 9);
vector<float>descriptors;//结果数组
hog->compute(trainTempImg, descriptors, cv::Size(1, 1), cv::Size(0, 0));
//cout << "HOG dims: " << descriptors.size() << endl;
cv::Mat SVMtrainMat(1, descriptors.size(), CV_32FC1);
int n = 0;
for (vector<float>::iterator iter = descriptors.begin(); iter != descriptors.end(); iter++)
{
SVMtrainMat.at<float>(0, n) = *iter;
n++;
}
int ret = svm->predict(SVMtrainMat);//检测结果
cout << ret << endl;
cv::Mat svm_result;
svm->predict(SVMtrainMat, svm_result, cv::ml::StatModel::Flags::RAW_OUTPUT);
float dist = svm_result.at<float>(0, 0);
float confidence = (1.0 / (1.0 + exp(-dist)));
return 0;
}