OpenCV 神经网络

简要介绍

OpenCV的人工神经网络是机器学习算法中的其中一种，使用的是多层感知器（Multi- Layer Perception，MLP），是常见的一种ANN算法。MLP算法一般包括三层，分别是一个输入层，一个输出层和一个或多个隐藏层的神经网络组成。每一层由一个或多个神经元互相连结。一个“神经元”的输出就可以是另一个“神经元”的输入。例如，下图是一个简单3层的神经元感知器：（3个输入，2个输出以及包含5个神经元的隐藏层）

MLP算法中，每个神经元都有几个输入和输出神经元，每个神经元通过输入权重加上偏置计算输出值，并选择一种激励函数进行转换。

激励函数常见的有三种，分别是恒等函数，Sigmoid函数和高斯函数，OpenCV默认的是Sigmoid函数。Sigmoid函数的公式

下图是Sigmoid函数的alpha参数和Beta参数为1的图像。

OpenCV中的MLP

OpenCV中的ANN算法通过训练来计算和学习更新每一层的权重，突触以及神经元。为了训练出分类器，需要创建两个数据矩阵，一个是特征数据矩阵，一个是标签矩阵。但要注意的是标签矩阵是一个N*M的矩阵，N表示训练样本数，M是类标签。如果第i行的样本属于第j类，那么该标签矩阵的（i，j）位置为1。
OpenCV中ANN定义了CvANN_MLP类。使用ANN算法之前，必须先初始化参数，比如神经网络的层数、神经元数，激励函数、α和β。然后使用train函数进行训练，训练完成可以训练好的参数以xml的格式保存在本地文件夹。最后就可以使用predict函数来预测识别。

OpenCV ANN的数字识别的例子

#include "stdafx.h"
#include <opencv2/opencv.hpp>
#include <iostream>
#include <fstream>
#include <sstream>
#include <math.h>
#include <vector>
#include <windows.h>
#include <io.h>
#include <time.h>

using namespace cv;
using namespace std;


#define HORIZONTAL    1
#define VERTICAL    0

CvANN_MLP  ann;

const char strCharacters[] = { ‘0‘,‘1‘,‘2‘,‘3‘,‘4‘,‘5‘,‘6‘,‘7‘,‘8‘,‘9‘};
const int numCharacter = 10; 

const int numNeurons = 40;
const int predictSize = 10;


void generateRandom(int n, int test_num, int min, int max, vector<int>*mark_samples)
{
    int range = max - min;
    int index = rand() % range + min;
    if (mark_samples->at(index) == 0)
    {
        mark_samples->at(index) = 1;
        n++;
    }

    if (n < test_num)
        generateRandom(n, test_num, min, max, mark_samples);

}


vector<string> getFiles(const string &folder,
    const bool all /* = true */) {
    vector<string> files;
    list<string> subfolders;
    subfolders.push_back(folder);

    while (!subfolders.empty()) {
        string current_folder(subfolders.back());

        if (*(current_folder.end() - 1) != ‘/‘) {
            current_folder.append("/*");
        }
        else {
            current_folder.append("*");
        }

        subfolders.pop_back();

        struct _finddata_t file_info;
        long file_handler = _findfirst(current_folder.c_str(), &file_info);

        while (file_handler != -1) {
            if (all &&
                (!strcmp(file_info.name, ".") || !strcmp(file_info.name, ".."))) {
                if (_findnext(file_handler, &file_info) != 0) break;
                continue;
            }

            if (file_info.attrib & _A_SUBDIR) {
                // it‘s a sub folder
                if (all) {
                    // will search sub folder
                    string folder(current_folder);
                    folder.pop_back();
                    folder.append(file_info.name);

                    subfolders.push_back(folder.c_str());
                }
            }
            else {
                // it‘s a file
                string file_path;
                // current_folder.pop_back();
                file_path.assign(current_folder.c_str()).pop_back();
                file_path.append(file_info.name);

                files.push_back(file_path);
            }

            if (_findnext(file_handler, &file_info) != 0) break;
        }  // while
        _findclose(file_handler);
    }

    return files;
}

void AppendText(string filename, string text)
{
    fstream ftxt;
    ftxt.open(filename, ios::out | ios::app);
    if (ftxt.fail())
    {
        cout << "创建文件失败!" << endl;
        getchar();
    }
    ftxt << text << endl;
    ftxt.close();
}

// ！获取垂直和水平方向直方图
Mat ProjectedHistogram(Mat img, int t)
{
    int sz = (t) ? img.rows : img.cols;
    Mat mhist = Mat::zeros(1, sz, CV_32F);

    for (int j = 0; j<sz; j++) {
        Mat data = (t) ? img.row(j) : img.col(j);
        mhist.at<float>(j) = countNonZero(data);    //统计这一行或一列中，非零元素的个数，并保存到mhist中
    }

    //Normalize histogram
    double min, max;
    minMaxLoc(mhist, &min, &max);

    if (max>0)
        mhist.convertTo(mhist, -1, 1.0f / max, 0);//用mhist直方图中的最大值，归一化直方图

    return mhist;
}

Mat features(Mat in, int sizeData)
{
    //Histogram features

    Mat vhist = ProjectedHistogram(in, VERTICAL);
    Mat hhist = ProjectedHistogram(in, HORIZONTAL);
    //Low data feature
    Mat lowData;
    resize(in, lowData, Size(sizeData, sizeData));


    //Last 10 is the number of moments components
    int numCols = vhist.cols + hhist.cols + lowData.cols*lowData.cols;
    //int numCols = vhist.cols + hhist.cols;
    Mat out = Mat::zeros(1, numCols, CV_32F);
    //Asign values to feature,ANN的样本特征为水平、垂直直方图和低分辨率图像所组成的矢量
    int j = 0;
    for (int i = 0; i<vhist.cols; i++)
    {
        out.at<float>(j) = vhist.at<float>(i);
        j++;
    }
    for (int i = 0; i<hhist.cols; i++)
    {
        out.at<float>(j) = hhist.at<float>(i);
        j++;
    }
    for (int x = 0; x<lowData.cols; x++)
    {
        for (int y = 0; y<lowData.rows; y++) {
            out.at<float>(j) = (float)lowData.at<unsigned char>(x, y);
            j++;
        }
    }
    //if(DEBUG)
    //  cout << out << "\n===========================================\n";
    return out;
}

void annTrain(Mat TrainData, Mat classes, int nNeruns)
{
    ann.clear();
    Mat layers(1, 3, CV_32SC1);
    layers.at<int>(0) = TrainData.cols;
    layers.at<int>(1) = nNeruns;
    layers.at<int>(2) = numCharacter;
    ann.create(layers, CvANN_MLP::SIGMOID_SYM, 1, 1);

    //Prepare trainClases
    //Create a mat with n trained data by m classes
    Mat trainClasses;
    trainClasses.create(TrainData.rows, numCharacter, CV_32FC1);
    for (int i = 0; i < trainClasses.rows; i++)
    {
        for (int k = 0; k < trainClasses.cols; k++)
        {
            //If class of data i is same than a k class
            if (k == classes.at<int>(i))
                trainClasses.at<float>(i, k) = 1;
            else
                trainClasses.at<float>(i, k) = 0;
        }
    }
    Mat weights(1, TrainData.rows, CV_32FC1, Scalar::all(1));

    //Learn classifier
    // ann.train( TrainData, trainClasses, weights );

    //Setup the BPNetwork

    // Set up BPNetwork‘s parameters
    CvANN_MLP_TrainParams params;
    params.train_method = CvANN_MLP_TrainParams::BACKPROP;
    params.bp_dw_scale = 0.1;
    params.bp_moment_scale = 0.1;

    //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.;

    ann.train(TrainData, trainClasses, Mat(), Mat(), params);

}

int recog(Mat features)
{
    int result = -1;
    Mat Predict_result(1, numCharacter, CV_32FC1);
    ann.predict(features, Predict_result);
    Point maxLoc;
    double maxVal;

    minMaxLoc(Predict_result, 0, &maxVal, 0, &maxLoc);

    return maxLoc.x;
}

float ANN_test(Mat samples_set, Mat sample_labels)
{
    int correctNum = 0;
    float accurate = 0;
    for (int i = 0; i < samples_set.rows; i++)
    {
        int result = recog(samples_set.row(i));
        if (result == sample_labels.at<int>(i))
            correctNum++;
    }
    accurate = (float)correctNum / samples_set.rows;
    return accurate;
}

int saveTrainData()
{
    cout << "Begin saveTrainData" << endl;
    Mat classes;
    Mat trainingDataf5;
    Mat trainingDataf10;
    Mat trainingDataf15;
    Mat trainingDataf20;

    vector<int> trainingLabels;
    string path = "charSamples";

    for (int i = 0; i < numCharacter; i++)
    {
        cout << "Character: " << strCharacters[i] << "\n";
        stringstream ss(stringstream::in | stringstream::out);
        ss << path << "/" << strCharacters[i];

        auto files = getFiles(ss.str(),1);

        int size = files.size();
        for (int j = 0; j < size; j++)
        {
            cout << files[j].c_str() << endl;
            Mat img = imread(files[j].c_str(), 0);
            Mat f5 = features(img, 5);
            Mat f10 = features(img, 10);
            Mat f15 = features(img, 15);
            Mat f20 = features(img, 20);

            trainingDataf5.push_back(f5);
            trainingDataf10.push_back(f10);
            trainingDataf15.push_back(f15);
            trainingDataf20.push_back(f20);
            trainingLabels.push_back(i);            //每一幅字符图片所对应的字符类别索引下标
        }
    }



    trainingDataf5.convertTo(trainingDataf5, CV_32FC1);
    trainingDataf10.convertTo(trainingDataf10, CV_32FC1);
    trainingDataf15.convertTo(trainingDataf15, CV_32FC1);
    trainingDataf20.convertTo(trainingDataf20, CV_32FC1);
    Mat(trainingLabels).copyTo(classes);

    FileStorage fs("train/features_data.xml", FileStorage::WRITE);
    fs << "TrainingDataF5" << trainingDataf5;
    fs << "TrainingDataF10" << trainingDataf10;
    fs << "TrainingDataF15" << trainingDataf15;
    fs << "TrainingDataF20" << trainingDataf20;
    fs << "classes" << classes;
    fs.release();

    cout << "End saveTrainData" << endl;

    return 0;
}


void ANN_Cross_Train_and_Test(int Imagsize, int Layers )
{

    String training;
    Mat TrainingData;
    Mat Classes;

    FileStorage fs;
    fs.open("train/features_data.xml", FileStorage::READ);


    cout << "Begin to ANN_Cross_Train_and_Test " << endl;

    char *txt = new char[50];
    sprintf(txt, "交叉训练，特征维度%d,网络层数%d", 40 + Imagsize * Imagsize, Layers);
    AppendText("output.txt", txt);
    cout << txt << endl;
    stringstream ss(stringstream::in | stringstream::out);
    ss << "TrainingDataF" << Imagsize;
    training = ss.str();

    fs[training] >> TrainingData;
    fs["classes"] >> Classes;
    fs.release();

    float result = 0.0;

    srand(time(NULL));

    vector<int> markSample(TrainingData.rows, 0);

    generateRandom(0, 100, 0, TrainingData.rows - 1, &markSample);

    Mat train_set, train_labels;
    Mat sample_set, sample_labels;

    for (int i = 0; i < TrainingData.rows; i++)
    {
        if (markSample[i] == 1)
        {
            sample_set.push_back(TrainingData.row(i));
            sample_labels.push_back(Classes.row(i));
        }
        else
        {
            train_set.push_back(TrainingData.row(i));
            train_labels.push_back(Classes.row(i));
        }
    }

    annTrain(train_set, train_labels, Layers);

    result = ANN_test(sample_set, sample_labels);


    sprintf(txt, "正确率%f\n", result);
    cout << txt << endl;
    AppendText("output.txt", txt);

    cout << "End the ANN_Cross_Train_and_Test" << endl;

    cout << endl;


}


void ANN_test_Main()
{
    int DigitSize[4] = { 5, 10, 15, 20};
    int LayerNum[14] = { 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 120, 150, 200, 500 };
    for (int i = 0; i < 4; i++)
    {
        for (int j = 0; j < 14; j++)
        {
            ANN_Cross_Train_and_Test(DigitSize[i], LayerNum[j]);
        }
    }
}


void ANN_saveModel(int _predictsize, int _neurons)
{
    FileStorage fs;
    fs.open("train/features_data.xml", FileStorage::READ);

    Mat TrainingData;
    Mat Classes;

    string training;
    if (1)
    {
        stringstream ss(stringstream::in | stringstream::out);
        ss << "TrainingDataF" << _predictsize;
        training = ss.str();
    }

    fs[training] >> TrainingData;
    fs["classes"] >> Classes;

    //train the Ann
    cout << "Begin to saveModelChar predictSize:" << _predictsize
        << " neurons:" << _neurons << endl;


    annTrain(TrainingData, Classes, _neurons);



    cout << "End the saveModelChar" << endl;


    string model_name = "train/ann10_40.xml";
    //if(1)
    //{
    //  stringstream ss(stringstream::in | stringstream::out);
    //  ss << "ann_prd" << _predictsize << "_neu"<< _neurons << ".xml";
    //  model_name = ss.str();
    //}

    FileStorage fsTo(model_name, cv::FileStorage::WRITE);
    ann.write(*fsTo, "ann");
}

int main()
{
    cout << "To be begin." << endl;

    saveTrainData();

    //ANN_saveModel(10, 40);

    ANN_test_Main();





    cout << "To be end." << endl;
    int end;
    cin >> end;
    return 0;
}

程序运行的结果为：

交叉训练，特征维度65,网络层数10
正确率1.000000

交叉训练，特征维度65,网络层数20
正确率1.000000

交叉训练，特征维度65,网络层数30
正确率1.000000

交叉训练，特征维度65,网络层数40
正确率0.990000

交叉训练，特征维度65,网络层数50
正确率0.990000

交叉训练，特征维度65,网络层数60
正确率0.990000

交叉训练，特征维度65,网络层数70
正确率0.980000

交叉训练，特征维度65,网络层数80
正确率0.990000

交叉训练，特征维度65,网络层数90
正确率1.000000

交叉训练，特征维度65,网络层数100
正确率1.000000

交叉训练，特征维度65,网络层数120
正确率0.990000

交叉训练，特征维度65,网络层数150
正确率1.000000

交叉训练，特征维度65,网络层数200
正确率1.000000

交叉训练，特征维度65,网络层数500
正确率0.870000

交叉训练，特征维度140,网络层数10
正确率0.990000

交叉训练，特征维度140,网络层数20
正确率1.000000

交叉训练，特征维度140,网络层数30
正确率1.000000

交叉训练，特征维度140,网络层数40
正确率1.000000

交叉训练，特征维度140,网络层数50
正确率0.990000

交叉训练，特征维度140,网络层数60
正确率1.000000

交叉训练，特征维度140,网络层数70
正确率1.000000

交叉训练，特征维度140,网络层数80
正确率0.920000

交叉训练，特征维度140,网络层数90
正确率0.920000

交叉训练，特征维度140,网络层数100
正确率0.850000

交叉训练，特征维度140,网络层数120
正确率1.000000

交叉训练，特征维度140,网络层数150
正确率0.750000

交叉训练，特征维度140,网络层数200
正确率0.990000

交叉训练，特征维度140,网络层数500
正确率0.880000

交叉训练，特征维度265,网络层数10
正确率1.000000

交叉训练，特征维度265,网络层数20
正确率0.990000

交叉训练，特征维度265,网络层数30
正确率1.000000

交叉训练，特征维度265,网络层数40
正确率1.000000

交叉训练，特征维度265,网络层数50
正确率1.000000

交叉训练，特征维度265,网络层数60
正确率1.000000

交叉训练，特征维度265,网络层数70
正确率1.000000

交叉训练，特征维度265,网络层数80
正确率1.000000

交叉训练，特征维度265,网络层数90
正确率1.000000

交叉训练，特征维度265,网络层数100
正确率0.920000

交叉训练，特征维度265,网络层数120
正确率1.000000

交叉训练，特征维度265,网络层数150
正确率1.000000

交叉训练，特征维度265,网络层数200
正确率0.990000

交叉训练，特征维度265,网络层数500
正确率0.840000

交叉训练，特征维度440,网络层数10
正确率1.000000

交叉训练，特征维度440,网络层数20
正确率1.000000

交叉训练，特征维度440,网络层数30
正确率0.980000

交叉训练，特征维度440,网络层数40
正确率1.000000

交叉训练，特征维度440,网络层数50
正确率0.990000

交叉训练，特征维度440,网络层数60
正确率0.930000

交叉训练，特征维度440,网络层数70
正确率1.000000

交叉训练，特征维度440,网络层数80
正确率1.000000

交叉训练，特征维度440,网络层数90
正确率0.900000

交叉训练，特征维度440,网络层数100
正确率1.000000

交叉训练，特征维度440,网络层数120
正确率1.000000

交叉训练，特征维度440,网络层数150
正确率0.880000

交叉训练，特征维度440,网络层数200
正确率0.840000

交叉训练，特征维度440,网络层数500
正确率0.830000

参考资料

Neural Networks — OpenCV 3.0.0-dev documentation
http://docs.opencv.org/3.0-beta/modules/ml/doc/neural_networks.html
UFLDL教程 - Ufldl
http://ufldl.stanford.edu/wiki/index.php/UFLDL%E6%95%99%E7%A8%8B
从线性分类器到卷积神经网络 // 在路上
http://zhangliliang.com/2014/06/14/from-lr-to-cnn/
神经网络研究项目–以工程师的视角 - jsxyhelu - 博客园
http://www.cnblogs.com/jsxyhelu/p/4306753.html
OCR using Artificial Neural Network (OpenCV) – Part 1 | Nithin Raj S.
http://www.nithinrajs.in/ocr-using-artificial-neural-network-opencv-part-1/
liuruoze/EasyPR
https://github.com/liuruoze/EasyPR
willhope/code
https://github.com/willhope/code
OpenCV进阶之路：神经网络识别车牌字符 - ☆Ronny丶 - 博客园
http://www.cnblogs.com/ronny/p/opencv_road_more_01.html#commentform
感知器与梯度下降 - ☆Ronny丶 - 博客园
http://www.cnblogs.com/ronny/p/ann_01.html
神经网络入门(连载之一) - zzwu的专栏 - 博客频道 - CSDN.NET
http://blog.csdn.net/zzwu/article/details/574931/