OpenCV——手势识别

Posted 2020-08-17 Not-Bad

使用ANN神经网络训练数据后进行手势识别。

#include "header.h"


int main()
{
    const int sample_num = 10;              //训练每类图片数量
    const int class_num = 3;                //训练类数3:石头剪刀布
    const int image_cols = 30;
    const int image_rows = 30;
    string  Name, Path;

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

    cout << "training Data.........\\n";
    for (int i = 0; i < class_num; i++){

        int j = 0;
        Path = getstring(i + 1) + "/" + "*.*";

        HANDLE hFile;
        LPCTSTR lp = StringToWchar(Path);
        WIN32_FIND_DATA pNextInfo;
        hFile = FindFirstFile(lp, &pNextInfo);
        if (hFile == INVALID_HANDLE_VALUE){
            cout << "failed" << endl;
            exit(-1);//搜索失败
        }
        cout << "folder name:" << i + 1 << endl;
        
        do{
            //必须加这句，不然会加载.和..的文件而加载不了图片，
            if (pNextInfo.cFileName[0] == \'.\')continue;   

            cout << "file name" << WcharToChar(pNextInfo.cFileName) << endl;
            Mat srcImage = imread(getstring(i+1) + "/" + WcharToChar(pNextInfo.cFileName), CV_LOAD_IMAGE_GRAYSCALE);
            Mat trainImage;
            
            //if (!srcImage.empty())cout << " done \\n";
            //处理样本图像
            resize(srcImage, trainImage, Size(image_cols, image_rows), (0, 0), (0, 0), CV_INTER_AREA); 
            Canny(trainImage, trainImage, 150, 100, 3, false);
            for (int k = 0; k < image_rows * image_cols; k++){
                //cout << "矩阵 k-- " << k << "   j--" << j << "   i--" << i << endl;
                trainingData[i*sample_num + j][k] = (float)trainImage.data[k];

            }
            j++;
        } while (FindNextFile(hFile, &pNextInfo));
    }

    // 训练好的矩阵
    Mat DataMat(class_num*sample_num, image_rows*image_cols, CV_32FC1, trainingData);
    cout << "DataMat   done~" << endl;

    // 初始化标签 
    // 0-石头 1-剪刀 2-布
    for (int i = 0; i < class_num ; i++){
        for (int j = 0; j < sample_num; j++){
            for (int k = 0; k < class_num; k++){
                if (k == i)labels[i*sample_num + j][k] = 1;
                else labels[i*sample_num + j][k] = 0;
            }
        }
    }

    // 标签矩阵
    Mat labelsMat(class_num*sample_num, class_num, CV_32FC1, labels);
    cout << "labelsMat  done~" << endl;

    //训练代码
    CvANN_MLP bp;
    CvANN_MLP_TrainParams params;
    params.train_method = CvANN_MLP_TrainParams::BACKPROP;
    params.bp_dw_scale = 0.001;    
    params.bp_moment_scale = 0.1;
    //cvTermCriteria 迭代终止规则
    params.term_crit = cvTermCriteria(CV_TERMCRIT_ITER | CV_TERMCRIT_EPS, 10000, 0.0001);     

    //设置网络层数
    Mat layerSizes = (Mat_<int>(1, 4) << image_rows*image_cols, int(image_rows*image_cols / 2), 
                                                                 int(image_rows*image_cols / 2), class_num);
    bp.create(layerSizes, CvANN_MLP::SIGMOID_SYM, 1.0, 1.0);     
    cout << "training...." << endl;
    bp.train(DataMat, labelsMat, Mat(), Mat(), params);

    bp.save("detect_gesture.xml"); 
    cout << "done" << endl;

    //测试神经网络
    cout << "testing...." << endl;

    Mat test = imread("test.jpg");   
    Mat temp;
    resize(test, temp, Size(image_cols, image_rows), (0, 0), (0, 0), CV_INTER_AREA);
    Canny(temp, temp, 150, 100, 3, false);
    Mat_<float>sample(1, image_rows*image_cols);
    for (int i = 0; i<image_rows*image_cols; ++i){
        sample.at<float>(0, i) = (float)temp.data[i];
    }

    Mat result;
    bp.predict(sample, result);

    float* p = result.ptr<float>(0);
    float max = -1, min = 0;
    int index = 0;
    for (int i = 0; i<class_num; i++)
    {
        cout << (float)(*(p + i)) << " ";
        if (i == class_num - 1)
            cout << endl;
        if ((float)(*(p + i))>max)
        {
            min = max;
            max = (float)(*(p + i));
            index = i;
        }
        else
        {
            if (min < (float)(*(p + i)))
                min = (float)(*(p + i));
        }
    }
    cout << "Your choice :" << choice[index] << endl << "识别率:" 
                                              << (((max - min) * 100) > 100 ? 100 : ((max - min) * 100)) << endl;

    
    //石头剪刀布——游戏开局~
    int computer = random(3);
    cout << "computer\'s choice :" << choice[computer] << endl;
    if (computer == index) cout << "A Draw   -_- " << endl << endl;
    else if ((computer < index && (index - computer == 1)) || (computer == 2 && index == 0)){
        cout << "You Lose  T_T " << endl << endl;
    }
    else cout << "You Win o   *￣▽￣* " << endl << endl;

    system("pause");
    waitKey(100);
    return 0;
}

 运行一次后，不用每次都训练数据，直接加载第一次保存的 "detect_gesture.xml"即可

CvANN_MLP bp;
CvANN_MLP_TrainParams params;
bp.load("detect_gesture.xml");

PS:

//CvTermCriteria()
//迭代算法的终止准则
#define CV_TERMCRIT_ITER    1
#define CV_TERMCRIT_NUMBER  CV_TERMCRIT_ITER
#define CV_TERMCRIT_EPS     2

typedef struct CvTermCriteria
 {
  int    type;                      // CV_TERMCRIT_ITER 和CV_TERMCRIT_EPS二值之一，或者二者的组合 
  int    max_iter;                  // 最大迭代次数 
  double epsilon;                   // 结果的精确性 
 }
 CvTermCriteria;
// 构造函数 
inline  CvTermCriteria  cvTermCriteria( int type, int max_iter, double epsilon );
// 在满足max_iter和epsilon的条件下检查终止准则并将其转换使得type=CV_TERMCRIT_ITER+CV_TERMCRIT_EPS 
CvTermCriteria cvCheckTermCriteria( CvTermCriteria criteria, double default_eps, int default_max_iters );

 文件查找相关