opencv 车牌定位及分割

车牌识别大概步骤可分为：车牌定位，字符分割，字符识别三个步骤。

细分点可以有以下几个步骤：

（1）将图片灰度化与二值化

（2）去噪，然后切割成一个一个的字符

（3）提取每一个字符的特征，生成特征矢量或特征矩阵

（4）分类与学习。将特征矢量或特征矩阵与样本库进行比对，挑选出相似的那类样本，将这类样本的值作为输出结果。

下面是车牌识别的第一个步骤，opencv源代码中sample有一个识别矩形的例子，网上资料说改改此代码就可以定位车牌，没有验证过，先贴个代码，权当记录一下，有时间的话再去实践一下。

也可参考以下文章：

http://blog.csdn.net/hbclc/archive/2007/10/14/1824365.aspx

代码如下：

//
// The full "Square Detector" program.
// It loads several images subsequentally and tries to find squares in
// each image
//
#ifdef _CH_
#pragma package <opencv>
#endif

#define CV_NO_BACKWARD_COMPATIBILITY

#include "cv.h"
#include "highgui.h"
#include <stdio.h>
#include <math.h>
#include <string.h>

int thresh = 50;
IplImage* img = 0;
IplImage* img0 = 0;
CvMemStorage* storage = 0;
const char* wndname = "Square Detection Demo";

// helper function:
// finds a cosine of angle between vectors
// from pt0->pt1 and from pt0->pt2
double angle( CvPoint* pt1, CvPoint* pt2, CvPoint* pt0 )
{
    double dx1 = pt1->x - pt0->x;
    double dy1 = pt1->y - pt0->y;
    double dx2 = pt2->x - pt0->x;
    double dy2 = pt2->y - pt0->y;
    return (dx1*dx2 + dy1*dy2)/sqrt((dx1*dx1 + dy1*dy1)*(dx2*dx2 + dy2*dy2) + 1e-10);
}

// returns sequence of squares detected on the image.
// the sequence is stored in the specified memory storage
CvSeq* findSquares4( IplImage* img, CvMemStorage* storage )
{
    CvSeq* contours;
    int i, c, l, N = 11;
    CvSize sz = cvSize( img->width & -2, img->height & -2 );    //保证最后一位是偶数,by sing 2010-10-11
    IplImage* timg = cvCloneImage( img ); // make a copy of input image
    IplImage* gray = cvCreateImage( sz, 8, 1 );
    IplImage* pyr = cvCreateImage( cvSize(sz.width/2, sz.height/2), 8, 3 );
    IplImage* tgray;
    CvSeq* result;
    double s, t;
    // create empty sequence that will contain points -
    // 4 points per square (the square's vertices)
    CvSeq* squares = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvPoint), storage );

    // select the maximum ROI in the image
    // with the width and height divisible by 2
    cvSetImageROI( timg, cvRect( 0, 0, sz.width, sz.height ));

    // down-scale and upscale the image to filter out the noise
    cvPyrDown( timg, pyr, 7 );
    cvPyrUp( pyr, timg, 7 );
    tgray = cvCreateImage( sz, 8, 1 );

    // find squares in every color plane of the image
    for( c = 0; c < 3; c++ )
    {
        // extract the c-th color plane
        cvSetImageCOI( timg, c+1 );
        cvCopy( timg, tgray, 0 );

        // try several threshold levels
        for( l = 0; l < N; l++ )
        {
            // hack: use Canny instead of zero threshold level.
            // Canny helps to catch squares with gradient shading
            if( l == 0 )
            {
                // apply Canny. Take the upper threshold from slider
                // and set the lower to 0 (which forces edges merging)
                cvCanny( tgray, gray, 0, thresh, 5 );
                // dilate canny output to remove potential
                // holes between edge segments
                cvDilate( gray, gray, 0, 1 );
            }
            else
            {
                // apply threshold if l!=0:
                //     tgray(x,y) = gray(x,y) < (l+1)*255/N ? 255 : 0
                cvThreshold( tgray, gray, (l+1)*255/N, 255, CV_THRESH_BINARY );
            }

            // find contours and store them all as a list
            cvFindContours( gray, storage, &contours, sizeof(CvContour),
                CV_RETR_LIST, CV_CHAIN_APPROX_SIMPLE, cvPoint(0,0) );

            // test each contour
            while( contours )
            {
                // approximate contour with accuracy proportional
                // to the contour perimeter
                result = cvApproxPoly( contours, sizeof(CvContour), storage,
                    CV_POLY_APPROX_DP, cvContourPerimeter(contours)*0.02, 0 );
                // square contours should have 4 vertices after approximation
                // relatively large area (to filter out noisy contours)
                // and be convex.
                // Note: absolute value of an area is used because
                // area may be positive or negative - in accordance with the
                // contour orientation
                if( result->total == 4 &&
                    cvContourArea(result,CV_WHOLE_SEQ,0) > 1000 &&
                    cvCheckContourConvexity(result) )
                {
                    s = 0;

                    for( i = 0; i < 5; i++ )
                    {
                        // find minimum angle between joint
                        // edges (maximum of cosine)
                        if( i >= 2 )
                        {
                            t = fabs(angle(
                            (CvPoint*)cvGetSeqElem( result, i ),
                            (CvPoint*)cvGetSeqElem( result, i-2 ),
                            (CvPoint*)cvGetSeqElem( result, i-1 )));
                            s = s > t ? s : t;
                        }
                    }

                    // if cosines of all angles are small
                    // (all angles are ~90 degree) then write quandrange
                    // vertices to resultant sequence
                    if( s < 0.3 )
                        for( i = 0; i < 4; i++ )
                            cvSeqPush( squares,
                                (CvPoint*)cvGetSeqElem( result, i ));
                }

                // take the next contour
                contours = contours->h_next;
            }
        }
    }

    // release all the temporary images
    cvReleaseImage( &gray );
    cvReleaseImage( &pyr );
    cvReleaseImage( &tgray );
    cvReleaseImage( &timg );

    return squares;
}


// the function draws all the squares in the image
void drawSquares( IplImage* img, CvSeq* squares )
{
    CvSeqReader reader;
    IplImage* cpy = cvCloneImage( img );
    int i;

    // initialize reader of the sequence
    cvStartReadSeq( squares, &reader, 0 );

    // read 4 sequence elements at a time (all vertices of a square)
    for( i = 0; i < squares->total; i += 4 )
    {
        CvPoint pt[4], *rect = pt;
        int count = 4;

        // read 4 vertices
        CV_READ_SEQ_ELEM( pt[0], reader );
        CV_READ_SEQ_ELEM( pt[1], reader );
        CV_READ_SEQ_ELEM( pt[2], reader );
        CV_READ_SEQ_ELEM( pt[3], reader );

        // draw the square as a closed polyline
        cvPolyLine( cpy, &rect, &count, 1, 1, CV_RGB(0,255,0), 3, CV_AA, 0 );
    }

    // show the resultant image
    cvShowImage( wndname, cpy );
    cvReleaseImage( &cpy );
}


char* names[] = { "pic1.png", "pic2.png", "pic3.png",
                  "pic4.png", "pic5.png", "pic6.png", 0 };

int main(int argc, char** argv)
{
    int i, c;
    // create memory storage that will contain all the dynamic data
    storage = cvCreateMemStorage(0);

    for( i = 0; names[i] != 0; i++ )
    {
        // load i-th image
        img0 = cvLoadImage( names[i], 1 );
        if( !img0 )
        {
            printf("Couldn't load %s/n", names[i] );
            continue;
        }
        img = cvCloneImage( img0 );

        // create window and a trackbar (slider) with parent "image" and set callback
        // (the slider regulates upper threshold, passed to Canny edge detector)
        cvNamedWindow( wndname, 1 );

        // find and draw the squares
        drawSquares( img, findSquares4( img, storage ) );

        // wait for key.
        // Also the function cvWaitKey takes care of event processing
        c = cvWaitKey(0);
        // release both images
        cvReleaseImage( &img );
        cvReleaseImage( &img0 );
        // clear memory storage - reset free space position
        cvClearMemStorage( storage );
        if( (char)c == 27 )
            break;
    }

    cvDestroyWindow( wndname );

    return 0;
}

这几天研究了一下车牌字符分割的问题，前提是已经进行了车牌定位，角度校正等预处理。

用到的主要知识有：二值化，形态学操作，轮廓查找等。

字符分割网上资料比较少，本人接触opencv一段时间，自己瞎搞了一下，以此抛砖引玉，希望与各位交流一下。

以下为全部源代码：


//==============================================

//write by sing

//2010-10-10

//==============================================


#include "stdafx.h"

//找出含车牌文字的最左端
void findX(IplImage* img, int* min, int* max)
{
    int found = 0;
    CvScalar maxVal = cvRealScalar(img->width * 255);
    CvScalar val = cvRealScalar(0);
    CvMat data;
    int minCount = img->width * 255 / 5;
    int count = 0;

    for (int i = 0; i < img->width; i++) {
        cvGetCol(img, &data, i);
        val = cvSum(&data);
        if (val.val[0] < maxVal.val[0]) {
            count = val.val[0];
if (count > minCount && count < img->width * 255) {
                *max = i;
                if (found == 0) {
                    *min = i;
                    found = 1;
                }
            }
        }
    }

}

//找出含车牌文字的最上端，排除两颗螺丝的位置
void findY(IplImage* img, int* min, int* max)
{
    int found = 0;
    CvScalar maxVal = cvRealScalar(img->height * 255);
    CvScalar val = cvRealScalar(0);
    CvMat data;
    int minCount = img->width * 255 / 5;
    int count = 0;

    for (int i = 0; i < img->height; i++) {
        cvGetRow(img, &data, i);
        val = cvSum(&data);
        if (val.val[0] < maxVal.val[0]) {
            count = val.val[0];
if (count > minCount && count < img->height * 255) {
                *max = i;
                if (found == 0) {
                    *min = i;
                    found = 1;
                }
            }
        }
    }
}

//车牌字符的最小区域
CvRect findArea(IplImage* img)
{
    int minX, maxX;
    int minY, maxY;
  
    findX(img, &minX, &maxX);
    findY(img, &minY, &maxY);

    CvRect rc = cvRect(minX, minY, maxX - minX, maxY - minY);

    return rc;
}

int main(int argc, char* argv[])
{
    IplImage* imgSrc = cvLoadImage("cp.jpg", CV_LOAD_IMAGE_COLOR);
    IplImage* img_gray = cvCreateImage(cvGetSize(imgSrc), IPL_DEPTH_8U, 1);
  
    cvCvtColor(imgSrc, img_gray, CV_BGR2GRAY);
    cvThreshold(img_gray, img_gray, 100, 255, CV_THRESH_BINARY);

    //寻找最小区域,并截取
    CvRect rc = findArea(img_gray);
    cvSetImageROI(img_gray, rc);
    IplImage* img_gray2 = cvCreateImage(cvSize(rc.width, rc.height), IPL_DEPTH_8U, 1);
    cvCopyImage(img_gray, img_gray2);
    cvResetImageROI(img_gray);

    IplImage* imgSrc2 = cvCreateImage(cvSize(rc.width, rc.height), IPL_DEPTH_8U, 3);
    cvSetImageROI(imgSrc, rc);
    cvCopyImage(imgSrc, imgSrc2);
    cvResetImageROI(imgSrc);

    //形态学
    cvMorphologyEx(img_gray2, img_gray2, NULL, NULL, CV_MOP_CLOSE);

    CvSeq* contours = NULL;
    CvMemStorage* storage = cvCreateMemStorage(0);
    int count = cvFindContours(img_gray2, storage, &contours,
        sizeof(CvContour), CV_RETR_EXTERNAL);

    int idx = 0;
    char szName[56] = {0};

    for (CvSeq* c = contours; c != NULL; c = c->h_next) {

        //cvDrawContours(imgSrc2, c, CV_RGB(255, 0, 0), CV_RGB(255, 255, 0), 100);
        CvRect rc = cvBoundingRect(c);
        cvDrawRect(imgSrc2, cvPoint(rc.x, rc.y), cvPoint(rc.x + rc.width, rc.y + rc.height), CV_RGB(255, 0, 0));

        if (rc.width < imgSrc2->width / 10 && rc.height < imgSrc2->height / 5) {
            continue;
        }


        IplImage* imgNo = cvCreateImage(cvSize(rc.width, rc.height), IPL_DEPTH_8U, 3);
        cvSetImageROI(imgSrc2, rc);
        cvCopyImage(imgSrc2, imgNo);
        cvResetImageROI(imgSrc2);
      
        sprintf(szName, "wnd_%d", idx++);
        cvNamedWindow(szName);
        cvShowImage(szName, imgNo);
        cvReleaseImage(&imgNo);
    }
  
    cvNamedWindow("src");
    cvShowImage("src", imgSrc2);

    cvWaitKey(0);

    cvReleaseMemStorage(&storage);
    cvReleaseImage(&imgSrc);
    cvReleaseImage(&imgSrc2);
    cvReleaseImage(&img_gray);
    cvReleaseImage(&img_gray2);

    cvDestroyAllWindows();

    return 0;
}