【opencv】神经网络识别数字

2017/5/5 应人要求，添加了数字加字母的识别，见链接
http://blog.csdn.net/qq_15947787/article/details/72850445

文本仅对0-9这十个文件夹中sample_mun_perclass个样本进行训练，直接通过API函数FindFirstFile和FindNextFile得到目录下文件，不需要对图片名编号

用了一下午时间去调这个代码，所以还有很多不完善的地方，以后有时间再去完善，比如：
本文的测试图片仅仅是单张测试，如果要测试准确率，可以根据前面训练时批量读取图片的代码进行简单修改，即可进行批量测试。
参考链接：http://blog.csdn.net/qq_15947787/article/details/51384258

特征采用8*16的二值化图像构成的128维向量作为输入层，3层128维的隐藏层，10维的输出层。输出用{1,0,0,0,0,…0}{0,1,0,0,0,…0}{0,0,1,0,0,…0}……{0,0,0,0,0,…1}表示

测试图像来源：http://www.cnblogs.com/ronny/p/opencv_road_more_01.html

测试代码：

//opencv2.4.9 + vs2012 + 64位
#include <windows.h>
#include <iostream>
#include <opencv2/opencv.hpp>

using namespace cv;
using namespace std;

char* WcharToChar(const wchar_t* wp)  
{  
    char *m_char;
    int len= WideCharToMultiByte(CP_ACP,0,wp,wcslen(wp),NULL,0,NULL,NULL);  
    m_char=new char[len+1];  
    WideCharToMultiByte(CP_ACP,0,wp,wcslen(wp),m_char,len,NULL,NULL);  
    m_char[len]='\0';  
    return m_char;  
}  

wchar_t* CharToWchar(const char* c)  
{   
    wchar_t *m_wchar;
    int len = MultiByteToWideChar(CP_ACP,0,c,strlen(c),NULL,0);  
    m_wchar=new wchar_t[len+1];  
    MultiByteToWideChar(CP_ACP,0,c,strlen(c),m_wchar,len);  
    m_wchar[len]='\0';  
    return m_wchar;  
}  

wchar_t* StringToWchar(const string& s)  
{  
    const char* p=s.c_str();  
    return CharToWchar(p);  
}  

int main()
{
    const string fileform = "*.png";
    const string perfileReadPath = "charSamples";

    const int sample_mun_perclass = 20;//训练字符每类数量
    const int class_mun = 10;//训练字符类数

    const int image_cols = 8;
    const int image_rows = 16;
    string  fileReadName,
            fileReadPath;
    char temp[256];

    float trainingData[class_mun*sample_mun_perclass][image_rows*image_cols] = {{0}};//每一行一个训练样本
    float labels[class_mun*sample_mun_perclass][class_mun]={{0}};//训练样本标签

    for(int i=0;i<=class_mun-1;++i)//不同类
    {
        //读取每个类文件夹下所有图像
        int j = 0;//每一类读取图像个数计数
        sprintf(temp, "%d", i);
        fileReadPath = perfileReadPath + "/" + temp + "/" + fileform;
        cout<<"文件夹"<<i<<endl;
        HANDLE hFile;
        LPCTSTR lpFileName = StringToWchar(fileReadPath);//指定搜索目录和文件类型，如搜索d盘的音频文件可以是"D:\\*.mp3"
        WIN32_FIND_DATA pNextInfo;  //搜索得到的文件信息将储存在pNextInfo中;
        hFile = FindFirstFile(lpFileName,&pNextInfo);//请注意是 &pNextInfo , 不是 pNextInfo;
        if(hFile == INVALID_HANDLE_VALUE)
        {
            exit(-1);//搜索失败
        }
        //do-while循环读取
        do
        {   
            if(pNextInfo.cFileName[0] == '.')//过滤.和..
                continue;
            j++;//读取一张图
            //wcout<<pNextInfo.cFileName<<endl;
            printf("%s\n",WcharToChar(pNextInfo.cFileName));
            //对读入的图片进行处理
            Mat srcImage = imread( perfileReadPath + "/" + temp + "/" + WcharToChar(pNextInfo.cFileName),CV_LOAD_IMAGE_GRAYSCALE);
            Mat resizeImage;
            Mat trainImage;

            resize(srcImage,resizeImage,Size(image_cols,image_rows),(0,0),(0,0),CV_INTER_AREA);//使用象素关系重采样。当图像缩小时候，该方法可以避免波纹出现
            threshold(resizeImage,trainImage,0,255,CV_THRESH_BINARY|CV_THRESH_OTSU);

            for(int k = 0; k<image_rows*image_cols; ++k)
            {
                trainingData[i*sample_mun_perclass+(j-1)][k] = (float)trainImage.data[k];
                //trainingData[i*sample_mun_perclass+(j-1)][k] = (float)trainImage.at<unsigned char>((int)k/8,(int)k%8);//(float)train_image.data[k];
                //cout<<trainingData[i*sample_mun_perclass+(j-1)][k] <<" "<< (float)trainImage.at<unsigned char>(k/8,k%8)<<endl;
            }

        } while (FindNextFile(hFile,&pNextInfo) && j<sample_mun_perclass);//如果设置读入的图片数量，则以设置的为准，如果图片不够，则读取文件夹下所有图片

    }

    // Set up training data Mat
    Mat trainingDataMat(class_mun*sample_mun_perclass, image_rows*image_cols, CV_32FC1, trainingData);
    cout<<"trainingDataMat——OK！"<<endl;

    // Set up label data 
    for(int i=0;i<=class_mun-1;++i)
    {
        for(int j=0;j<=sample_mun_perclass-1;++j)
        {
            for(int k = 0;k<class_mun;++k)
            {
                if(k==i)
                    labels[i*sample_mun_perclass + j][k] = 1;
                else labels[i*sample_mun_perclass + j][k] = 0;
            }
        }
    }
    Mat labelsMat(class_mun*sample_mun_perclass, class_mun, CV_32FC1,labels);
    cout<<"labelsMat:"<<endl;
    cout<<labelsMat<<endl;
    cout<<"labelsMat——OK！"<<endl;

    //训练代码

    cout<<"training start...."<<endl;
    CvANN_MLP bp;
    // Set up BPNetwork's parameters
    CvANN_MLP_TrainParams params;
    params.train_method=CvANN_MLP_TrainParams::BACKPROP;
    params.bp_dw_scale=0.001;
    params.bp_moment_scale=0.1;
    params.term_crit = cvTermCriteria(CV_TERMCRIT_ITER|CV_TERMCRIT_EPS,10000,0.0001);  //设置结束条件
    //params.train_method=CvANN_MLP_TrainParams::RPROP;
    //params.rp_dw0 = 0.1;
    //params.rp_dw_plus = 1.2;
    //params.rp_dw_minus = 0.5;
    //params.rp_dw_min = FLT_EPSILON;
    //params.rp_dw_max = 50.;

    //Setup the BPNetwork
    Mat layerSizes=(Mat_<int>(1,5) << image_rows*image_cols,128,128,128,class_mun);
    bp.create(layerSizes,CvANN_MLP::SIGMOID_SYM,1.0,1.0);//CvANN_MLP::SIGMOID_SYM
                                               //CvANN_MLP::GAUSSIAN
                                               //CvANN_MLP::IDENTITY
    cout<<"training...."<<endl;
    bp.train(trainingDataMat, labelsMat, Mat(),Mat(), params);

    bp.save("../bpcharModel.xml"); //save classifier
    cout<<"training finish...bpModel1.xml saved "<<endl;


    //测试神经网络
    cout<<"测试："<<endl;
    Mat test_image = imread("test.png",CV_LOAD_IMAGE_GRAYSCALE);
    Mat test_temp;
    resize(test_image,test_temp,Size(image_cols,image_rows),(0,0),(0,0),CV_INTER_AREA);//使用象素关系重采样。当图像缩小时候，该方法可以避免波纹出现
    threshold(test_temp,test_temp,0,255,CV_THRESH_BINARY|CV_THRESH_OTSU);
    Mat_<float>sampleMat(1,image_rows*image_cols); 
    for(int i = 0; i<image_rows*image_cols; ++i)  
    {  
        sampleMat.at<float>(0,i) = (float)test_temp.at<uchar>(i/8,i%8);  
    }  

    Mat responseMat;  
    bp.predict(sampleMat,responseMat);  
    Point maxLoc;
    double maxVal = 0;
    minMaxLoc(responseMat,NULL,&maxVal,NULL,&maxLoc);
    cout<<"识别结果："<<maxLoc.x<<"  相似度:"<<maxVal*100<<"%"<<endl;
    imshow("test_image",test_image);  
    waitKey(0);

    return 0;
}

测试结果：

代码已打包上传：
http://download.csdn.net/detail/qq_15947787/9518680

5/13 优化trainingDataMat, labelsMat