opencv3/C++ ANN神经网络字母识别

调试平台：VS2015 windows10 64位

openc版本：opencv3.4.0

 

本例子采用opencv 自带的ANN神经网络识别算法，训练库是我自己用1280*720的摄像头采集的，该训练库包含了大写英文字母A~Z，共26个字母，每个字母1000张样本图片，图片尺寸才30*30。

大约花了3个多小时训练完成。

word中字体样本参数设置。

训练图片样本部分截图如下：

#include 
#include 
#include 
#include 
#include     

#include 
#include 
#include 

using namespace cv;
using namespace std;
using namespace ml;


#define CLASSSUM    26   // 图片共有26类
#define IMAGE_ROWS  30   // 统一图片高度
#define IMAGE_COLS  30   // 统一图片宽度
#define IMAGESSUM   1000   // 每一类图片张数

//std::string dirNum[CLASSSUM] = { "0", "1", "2", "3", "4", "5", "6", "7", "8", "9", "A", "B", "C", "D", "E", "F", "G", "H", "J", "K", "L", "M", "N", "P", "Q", "R", "S", "T", "U", "V", "W", "X", "Y", "Z" };
//std::string dirNum[CLASSSUM] = { "0", "1", "2", "3", "4", "5", "6", "7", "8", "9"};
std::string dirNum[CLASSSUM] = { "A", "B", "C", "D", "E", "F", "G", "H","I", "J", "K", "L", "M", "N", "O", "P", "Q", "R", "S", "T", "U", "V", "W", "X", "Y", "Z" };

float trainingData[CLASSSUM*IMAGESSUM][IMAGE_ROWS*IMAGE_COLS] = { { 0 } };  // 每一行一个训练图片
float labels[CLASSSUM*IMAGESSUM][CLASSSUM] = { { 0 } };                     // 训练样本标签

void TestXml()
{
    Ptr model = StatModel::load("./Image/English.xml");

    for (int dnum = 0; dnum < CLASSSUM; dnum++)
    {
        std::string inPath = ".\\Image\\english\\" + dirNum[dnum] + "\\*.jpg";
        intptr_t handle;
        struct _finddata_t fileinfo;
        handle = _findfirst(inPath.c_str(), &fileinfo);
        if (handle == -1) return;
        int num = 0;
        int imgok = 0;

        int imgnum = 0;
        do {
            std::string imgname = "./Image/english/" + dirNum[dnum] + "/" + fileinfo.name;
            //cout << imgname << endl;
            Mat srcImage = imread(imgname, 0);
            //imshow("srcImage", srcImage);
            //waitKey(1);
            if (!srcImage.empty())
            {
                //将测试图像转化为1*128的向量
                resize(srcImage, srcImage, Size(IMAGE_ROWS, IMAGE_COLS), (0, 0), (0, 0), INTER_AREA);
                threshold(srcImage, srcImage, 0, 255, CV_THRESH_BINARY | CV_THRESH_OTSU);
                Mat_ testMat(1, IMAGE_ROWS*IMAGE_COLS);
                for (int i = 0; i < IMAGE_ROWS*IMAGE_COLS; i++) {
                    testMat.at(0, i) = (float)srcImage.at(i / IMAGE_ROWS, i % IMAGE_COLS);
                }

                //使用训练好的MLP model预测测试图像
                Mat dst;
                model->predict(testMat, dst);
                //std::cout << "testMat: \n" << testMat << "\n" << std::endl;
                //std::cout << "dst: \n" << dst << "\n" << std::endl;
                double maxVal = 0;
                Point maxLoc;
                minMaxLoc(dst, NULL, &maxVal, NULL, &maxLoc);
                //std::cout  << dirNum[dnum] << " " << num++ << " 测试结果：" << dirNum[maxLoc.x] << "置信度:" << maxVal * 100 << "%" << std::endl;
                if (dirNum[dnum] == dirNum[maxLoc.x]) { 
                    imgok++; 
                }
                else
                {
                    std::string errPath = ".\\Image\\Error\\" + dirNum[dnum] + "\\" + fileinfo.name + " " + dirNum[maxLoc.x] + ".jpg";
                    // cout << errPath << endl;
                    imwrite(errPath, srcImage);
                }
                imgnum++;
            }
        } while (!_findnext(handle, &fileinfo) && imgnum < IMAGESSUM);

        std::cout << dirNum[dnum] << " " << imgok << endl;
        _findclose(handle);
    }
}

int main()
{
#if 0
	for (int i = 0; i < CLASSSUM; i++)
	{
		int k = 0;
		std::string inPath = ".\\Image\\english\\" + dirNum[i] + "\\*.jpg";
		intptr_t handle;
		struct _finddata_t fileinfo;
		handle = _findfirst(inPath.c_str(), &fileinfo);
		if (handle == -1) return -1;

		do {
			std::string imgname =  "./Image/english/" + dirNum[i] + "/" + fileinfo.name;
            cout << imgname << endl;
			Mat srcImage = imread(imgname, 0);
			if (srcImage.empty()) {
				std::cout << "Read image error:" << imgname << std::endl;
				return -1;
			}

            resize(srcImage, srcImage, Size(IMAGE_ROWS, IMAGE_COLS), (0, 0), (0, 0), INTER_AREA);
			threshold(srcImage, srcImage, 0, 255, CV_THRESH_BINARY | CV_THRESH_OTSU);

			for (int j = 0; jmodel = ANN_MLP::create();
	Mat layerSizes = (Mat_(1, 5) << IMAGE_ROWS*IMAGE_COLS, 128, 128, 128, CLASSSUM);
	model->setLayerSizes(layerSizes);
	model->setTrainMethod(ANN_MLP::BACKPROP, 0.001, 0.1);
	model->setActivationFunction(ANN_MLP::SIGMOID_SYM, 1.0, 1.0);
	model->setTermCriteria(TermCriteria(TermCriteria::MAX_ITER | TermCriteria::EPS, 10000, 0.0001));
	Ptr trainData = TrainData::create(trainingDataMat, ROW_SAMPLE, labelsMat);
	model->train(trainData);
	model->save("./Image/English.xml"); //保存训练结果
#endif
    // 测试训练结果
    cout << "------------------test-------------------" << endl;
    TestXml();
    cout << "------------------end-------------------" << endl;
    waitKey(0);
    while (1);
	return 0;
}

执行结果如下：

使用样本进行识别测试，26个字母，1000张样本，成功率100%。

实际测试效果演示如下，目前稍微测试了下，尚未发现识别错误的情况：

训练样本图片下载：

https://download.csdn.net/download/gs1069405343/11505319

源码下载

https://download.csdn.net/download/gs1069405343/11505454