车牌识别系统开发记录(三) 字符识别

秒速五厘米 2022-08-26 14:19 239阅读 0赞

这篇博文来谈谈车牌的字符识别。

目前，车牌字符识别算法主要是基于模板匹配、特征匹配或神经网络的方法。

在本文中我们主要来说说基于神经网络的字符识别方法，采用的是OpenCV中的CvANN\_MLP。关于神经网络的具体细节，可以参考我以前的博文：

[BP神经网络解析及Matlab实现][BP_Matlab]

更加细节的东西可以查看如下参考文献：

[Neural Networks【OpenCV Documentation】][Neural Networks_OpenCV Documentation]

[BackPropWikipedia【Wikipedia】][BackPropWikipedia_Wikipedia]

现在我们确定了字符识别的总体框架，那么先来说说字符的特征提取问题。这里我们主要考虑的是：

**垂直方向、水平方向的数据统计特征提取法：**

这种特征提取法就是自左向右对图像进行逐列的扫描，统计每列黑色像素的个数，然后，自上往下逐行扫描，统计每行的黑色像素的个数，将统计结果作为字符的特征向量，如果字符的宽度为w，高度为h，则特征向量维数是w+h。

![SouthEast][]

同时，为了能够提取更多的特征用于识别，还将输入字符每个点的值引入特征向量。

**Code：**

Mat OCR::features(Mat in, int sizeData){
        // 直方图特征，
        Mat vhist=ProjectedHistogram(in,VERTICAL);
        Mat hhist=ProjectedHistogram(in,HORIZONTAL);
        
        // 将输入字符resize为15*15大小
        Mat lowData;
        resize(in, lowData, Size(sizeData, sizeData) );
            
        // 特征向量维数
        int numCols=vhist.cols+hhist.cols+lowData.cols*lowData.cols;
        
        Mat out=Mat::zeros(1,numCols,CV_32F);
        
        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;
    }

Mat OCR::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);
        }
    
        //Normalize histogram
        double min, max;
        minMaxLoc(mhist, &min, &max);
        
        if(max>0)
            mhist.convertTo(mhist,-1 , 1.0f/max, 0);
    
        return mhist;
    }

不过，在上面的特征提取之前，我们其实还需要做几步预处理：

首先，就是**字符分割**。在字符分割里面，我们首先将输入车牌二值化，然后利用findContours寻找出每个字符的轮廓，再利用boundingRect定位出每个轮廓的矩形区域，然后割取出每个字符的区域。

**Code：**

vector<CharSegment> OCR::segment(Plate plate){
        
    	Mat input=plate.plateImg;
        vector<CharSegment> output;
    
        Mat img_threshold;
        threshold(input, img_threshold, 60, 255, CV_THRESH_BINARY_INV);
        
    	if(DEBUG)
            imshow("Threshold plate", img_threshold);
        
    	Mat img_contours;
        img_threshold.copyTo(img_contours);
    
        // 在车牌区域中寻找可能字符的的轮廓
        vector< vector< Point> > contours;
        findContours(img_contours,
                contours, 
                CV_RETR_EXTERNAL, 
                CV_CHAIN_APPROX_NONE); 
        
        
        cv::Mat result;
        img_threshold.copyTo(result);
        cvtColor(result, result, CV_GRAY2RGB);
        cv::drawContours(result,contours,
                -1, 
                cv::Scalar(255,0,0), 
                1); 
    
        vector<vector<Point> >::iterator itc= contours.begin();
           
        while (itc!=contours.end()) {
            
            Rect mr= boundingRect(Mat(*itc));
            rectangle(result, mr, Scalar(0,255,0));
    
            Mat auxRoi(img_threshold, mr);
            if(verifySizes(auxRoi)){
                auxRoi=preprocessChar(auxRoi);
                output.push_back(CharSegment(auxRoi, mr));
                rectangle(result, mr, Scalar(0,125,255));
            }
            ++itc;
        }
        if(DEBUG)
            cout << "Num chars: " << output.size() << "\n";
    
        if(DEBUG)
            imshow("SEgmented Chars", result);
    
        return output;
    }

效果：

![SouthEast 1][]

然后，就是对字符进行预处理（统一大小）。

**Code：**

// 这个函数主要是对输入图片归一化到统一的大小20×20
    Mat OCR::preprocessChar(Mat in){
        int h=in.rows;
        int w=in.cols;
    
        Mat transformMat=Mat::eye(2,3,CV_32F);
        int m=max(w,h);
        transformMat.at<float>(0,2)=m/2 - w/2;
        transformMat.at<float>(1,2)=m/2 - h/2;
    
        Mat warpImage(m,m, in.type());
        warpAffine(in, warpImage, transformMat, warpImage.size(), INTER_LINEAR, BORDER_CONSTANT, Scalar(0) );
    
        Mat out;
        resize(warpImage, out, Size(charSize, charSize) ); 
    
        return out;
    }

现在我们来介绍一下OpenCV中的神经网络方法。在OpenCV的documentation中有如下叙述：

the whole trained network works as follows:

1.  Take the feature vector as input. The vector size is equal to the size of the input layer.
2.  Pass values as input to the first hidden layer.
3.  Compute outputs of the hidden layer using the weights and the activation functions.
4.  Pass outputs further downstream until you compute the output layer.

So, to compute the network, you need to know all the weights  ![w^\{n+1)\}\_\{i,j\}][w_n_1_i_j]  . The weights are computed by the training algorithm. The algorithm takes a training set, multiple input vectors with the corresponding output vectors, and iteratively adjusts the weights to enable the network to give the desired response to the provided input vectors.

void OCR::train(Mat TrainData, Mat classes, int nlayers){
        Mat layers(1,3,CV_32SC1);
        layers.at<int>(0)= TrainData.cols;
        layers.at<int>(1)= nlayers;
        layers.at<int>(2)= numCharacters;
        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, numCharacters, 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 );
        trained=true;
    }
    
    int OCR::classify(Mat f){
        int result=-1;
        Mat output(1, numCharacters, CV_32FC1);
    
        ann.predict(f, output);
        Point maxLoc;
        double maxVal;
        minMaxLoc(output, 0, &maxVal, 0, &maxLoc);
        //We need know where in output is the max val, the x (cols) is the class.
    
        return maxLoc.x;
    }

本文地址：http://blog.csdn.net/linj\_m/article/details/23736381

更多图像处理、机器视觉资源请关注 博客： [LinJM-机器视觉][LinJM-] 微博：[林建民-机器视觉][-]

[BP_Matlab]: http://blog.csdn.net/linj_m/article/details/9897839
[Neural Networks_OpenCV Documentation]: http://docs.opencv.org/modules/ml/doc/neural_networks.html#cvann-mlp
[BackPropWikipedia_Wikipedia]: http://en.wikipedia.org/wiki/Backpropagation
[SouthEast]: /images/20220824/f4a6b6da535049dfabbe60fb1d86885e.png
[SouthEast 1]: /images/20220824/f093995a7f134f409cc034af66e58ca0.png
[w_n_1_i_j]: http://docs.opencv.org/_images/math/f6b53640af0c8e268dfc3b85ec667cd6cd89669b.png
[LinJM-]: http://blog.csdn.net/linj_m
[-]: http://weibo.com/cengineer