使用opencv的SVM和神经网络实现车牌识别(高质量的文章******)

本文转载自：http://blog.csdn.net/ap1005834/article/details/51340831

一、前言

本文参考自《深入理解OpenCV 实用计算机视觉项目解析》中的自动车牌识别项目，并对其中的方法理解后，再进行实践。深刻认识到实际上要完成车牌区域准确定位、车牌区域中字符的准确分割，字符准确识别这一系列步骤的困难。所以最后的识别效果也是有待进一步提高。

二、程序流程

程序流程如下所示：



相应的main函数如下

[cpp] view
plain copy

print?

#include "carID_Detection.h"

int main()

{

Mat img_input = imread("testCarID.jpg");

//如果读入图像失败

if(img_input.empty())

{

fprintf(stderr, "Can not load image %s\n", "testCarID.jpg");

return -1;

}

Mat hsvImg ;

cvtColor(img_input,hsvImg,CV_BGR2HSV);

vector<Mat> planes;

split(hsvImg,planes);

Mat sImg;

sImg = planes[1]; //获得红色分量

blur(sImg,sImg,Size(3,3)); //3*3高斯滤波

vector <RotatedRect> rects_sImg;

posDetect(sImg ,rects_sImg);

Mat grayImg;

RgbConvToGray(img_input ,grayImg);

medianBlur(grayImg,grayImg,3); //3*3中值滤波

vector <RotatedRect> rects_grayImg;

posDetect(grayImg ,rects_grayImg);

vector <RotatedRect> rects_closeImg; //车牌区域较为贴近

posDetect_closeImg(sImg ,rects_closeImg);

vector <RotatedRect> rects_optimal;

optimPosDetect(rects_sImg,rects_grayImg,rects_closeImg,rects_optimal);

vector <Mat> output_area;

normalPosArea(img_input ,rects_optimal,output_area); //获得144*33的候选车牌区域output_area

CvSVM svmClassifier;

svm_train(svmClassifier); //使用SVM对正负样本进行训练

vector<Mat> plates_svm; //需要把候选车牌区域output_area图像中每个像素点作为一行特征向量，后进行预测

for(int i=0;i< output_area.size(); ++i)

{

Mat img = output_area[i];

Mat p = img.reshape(1,1);

p.convertTo(p,CV_32FC1);

int response = (int)svmClassifier.predict( p );

if (response == 1)

plates_svm.push_back(output_area[i]); //保存预测结果

}

if(plates_svm.size() != 0)

{

imshow("Test", plates_svm[0]); //正确预测的话，就只有一个结果plates_svm[0]

waitKey(0);

}

else

{

std::cout<<"定位失败";

return -1;

}

//从SVM预测获得得车牌区域中分割得字符区域

vector <Mat> char_seg;

char_segment(plates_svm[0],char_seg);

//获得7个字符矩阵的相应特征矩阵

vector <Mat> char_feature;

char_feature.resize(7);

for (int i =0;i<char_seg.size() ;++ i)

features(char_seg[i], char_feature[i],5);

//神经网络训练

CvANN_MLP ann_classify;

ann_train(ann_classify, 34 ,48); //34为样本的类别数，48为隐藏层的神经元数

//字符预测

vector<int> char_result;

classify(ann_classify,char_feature,char_result);

//此函数等待按键，按键盘任意键就返回

svmClassifier.clear();

return 0;

}

三、代码简介
本文编程没有使用类的概念，纯属面向过程编程，以下为主要使用的函数

[cpp] view
plain copy

print?

//carID_Detection.h

#include "opencv2/core/core.hpp"

#include "opencv2//highgui/highgui.hpp"

#include "opencv2/imgproc/imgproc.hpp"

#include "opencv2/ml/ml.hpp"

#include <time.h>

#include <stdlib.h>

#include <iostream>

using namespace std;

using namespace cv;

void RgbConvToGray(const Mat& inputImage,Mat & outpuImage); //rgb转为灰度

void posDetect(Mat &, vector <RotatedRect> &); //粗步选取候选车牌区域

bool verifySizes(const RotatedRect & ); //车牌区域需要满足的条件

void posDetect_closeImg(Mat &inputImage , vector <RotatedRect> & rects ) ; //考虑到车牌距离非常近的时候的情况

bool verifySizes_closeImg(const RotatedRect & candidate); //距离近时的车牌区域需要满足的条件

void optimPosDetect(vector <RotatedRect> & rects_sImg , vector <RotatedRect> & rects_grayImg, //车牌区域进一步定位

vector <RotatedRect> & rects_closeImg,vector <RotatedRect> & rects_optimal );

float calOverlap(const Rect& box1,const Rect& box2); //计算2个矩阵的重叠比例

void normalPosArea(Mat &intputImg, vector<RotatedRect> &rects_optimal, vector <Mat>& output_area ); //车牌裁剪，标准化为144*33

void svm_train(CvSVM & ); //取出SVM.xml中的特征矩阵和标签矩阵进行训练

void char_segment(const Mat & inputImg,vector <Mat>&); //对车牌区域中的字符进行分割

bool char_verifySizes(const RotatedRect &); //字符区域需要满足的条件

void char_sort(vector <RotatedRect > & in_char ); //对字符区域进行排序

void features(const Mat & in , Mat & out ,int sizeData); //获得一个字符矩阵对应的特征向量

Mat projectHistogram(const Mat& img ,int t); //计算水平或累计直方图，取决于t为0还是1

void ann_train(CvANN_MLP &ann ,int numCharacters, int nlayers); //取出ann_xml中的数据，并进行神经网络训练

void classify(CvANN_MLP& ann, vector<Mat> &char_feature , vector<int> & char_result); //使用神经网络模型预测车牌字符，并打印至屏幕

[cpp] view
plain copy

print?

//carID_Detection.cpp

#include "carID_Detection.h"

void RgbConvToGray(const Mat& inputImage,Mat & outpuImage) //g = 0.3R+0.59G+0.11B

{

outpuImage = Mat(inputImage.rows ,inputImage.cols ,CV_8UC1);

for (int i = 0 ;i<inputImage.rows ;++ i)

{

uchar *ptrGray = outpuImage.ptr<uchar>(i);

const Vec3b * ptrRgb = inputImage.ptr<Vec3b>(i);

for (int j = 0 ;j<inputImage.cols ;++ j)

{

ptrGray[j] = 0.3*ptrRgb[j][2]+0.59*ptrRgb[j][1]+0.11*ptrRgb[j][0];

}

}

}

void posDetect_closeImg(Mat &inputImage , vector <RotatedRect> & rects ) //初步找到候选区域 rects

{

Mat img_canny;

Canny(inputImage, img_canny, 150, 220);

Mat img_threshold;

threshold(img_canny , img_threshold,0,255 , CV_THRESH_OTSU+CV_THRESH_BINARY); //otsu算法自动获得阈值

Mat element = getStructuringElement(MORPH_RECT ,Size(15 ,3)); //闭形态学的结构元素

morphologyEx(img_threshold ,img_threshold,CV_MOP_CLOSE,element); //形态学处理

//寻找车牌区域的轮廓

vector< vector <Point> > contours;

findContours(img_threshold ,contours,CV_RETR_EXTERNAL, CV_CHAIN_APPROX_NONE);//只检测外轮廓

//对候选的轮廓进行进一步筛选

vector< vector <Point> > ::iterator itc = contours.begin();

while( itc != contours.end())

{

RotatedRect mr = minAreaRect(Mat( *itc )); //返回每个轮廓的最小有界矩形区域

if(!verifySizes_closeImg(mr)) //判断矩形轮廓是否符合要求

{

itc = contours.erase(itc);

}

else

{

rects.push_back(mr);

++itc;

}

}

}

bool verifySizes_closeImg(const RotatedRect & candidate)

{

float error = 0.4;

const float aspect = 44/14; //长宽比

int min = 100*aspect*100; //最小区域

int max = 180*aspect*180; //最大区域

float rmin = aspect - aspect*error; //考虑误差后的最小长宽比

float rmax = aspect + aspect*error; //考虑误差后的最大长宽比

int area = candidate.size.height * candidate.size.width;

float r = (float)candidate.size.width/(float)candidate.size.height;

if(r <1)

r = 1/r;

if( (area < min || area > max) || (r< rmin || r > rmax) )

return false;

else

return true;

}

void posDetect(Mat &inputImage , vector <RotatedRect> & rects ) //初步找到候选区域 rects

{

Mat img_sobel;

Sobel(inputImage , img_sobel , CV_8U, 1,0,3,1,0);

Mat img_threshold;

threshold(img_sobel , img_threshold,0,255 , CV_THRESH_OTSU+CV_THRESH_BINARY); //otsu算法自动获得阈值

Mat element = getStructuringElement(MORPH_RECT ,Size(15 ,3)); //闭形态学的结构元素

morphologyEx(img_threshold ,img_threshold,CV_MOP_CLOSE,element);

//寻找车牌区域的轮廓

vector< vector <Point> > contours;

findContours(img_threshold ,contours,CV_RETR_EXTERNAL, CV_CHAIN_APPROX_NONE);//只检测外轮廓

//对候选的轮廓进行进一步筛选

vector< vector <Point> > ::iterator itc = contours.begin();

while( itc != contours.end())

{

RotatedRect mr = minAreaRect(Mat( *itc )); //返回每个轮廓的最小有界矩形区域

if(!verifySizes(mr)) //判断矩形轮廓是否符合要求

{

itc = contours.erase(itc);

}

else

{

rects.push_back(mr);

++itc;

}

}

}

bool verifySizes(const RotatedRect & candidate)

{

float error = 0.4;

const float aspect = 44/14; //长宽比

int min = 20*aspect*20; //最小区域

int max = 180*aspect*180; //最大区域

float rmin = aspect - 2*aspect*error; //考虑误差后的最小长宽比

float rmax = aspect + 2*aspect*error; //考虑误差后的最大长宽比

int area = candidate.size.height * candidate.size.width;

float r = (float)candidate.size.width/(float)candidate.size.height;

if(r <1)

r = 1/r;

if( (area < min || area > max) || (r< rmin || r > rmax) ) //满足该条件才认为该candidate为车牌区域

return false;

else

return true;

}

void optimPosDetect(vector <RotatedRect> & rects_sImg , vector <RotatedRect> & rects_grayImg,

vector <RotatedRect> & rects_closeImg,vector <RotatedRect> & rects_optimal )

{

for (int i=0;i<rects_sImg.size() ;++ i)

{

for (int j=0;j<rects_grayImg.size() ; ++j)

{

if (calOverlap(rects_sImg[i].boundingRect() , rects_grayImg[j].boundingRect()) > 0.2)

{

if(rects_sImg[i].boundingRect().width * rects_sImg[i].boundingRect().height

>= rects_grayImg[j].boundingRect().width * rects_grayImg[j].boundingRect().height)

rects_optimal.push_back(rects_sImg[i]);

else

rects_optimal.push_back(rects_grayImg[j]);

}

}

}

if (rects_closeImg.size()<2 ) //只考虑1个，为了速度

{

for (int i =0;i < rects_optimal.size();++ i )

for (int j =0;j < rects_closeImg.size();++ j)

{

if (( calOverlap(rects_optimal[i].boundingRect() , rects_closeImg[j].boundingRect()) < 0.2 &&

calOverlap(rects_optimal[i].boundingRect() , rects_closeImg[j].boundingRect()) > 0.05))

{

rects_optimal.push_back(rects_closeImg[j]);

}

}

}

}

float calOverlap(const Rect& box1,const Rect& box2)

{

if (box1.x > box2.x+box2.width) { return 0.0; }

if (box1.y > box2.y+box2.height) { return 0.0; }

if (box1.x+box1.width < box2.x) { return 0.0; }

if (box1.y+box1.height < box2.y) { return 0.0; }

float colInt = min(box1.x+box1.width,box2.x+box2.width) - max(box1.x, box2.x);

float rowInt = min(box1.y+box1.height,box2.y+box2.height) - max(box1.y,box2.y);

float intersection = colInt * rowInt;

float area1 = box1.width*box1.height;

float area2 = box2.width*box2.height;

return intersection / (area1 + area2 - intersection);

}

void normalPosArea(Mat &intputImg, vector<RotatedRect> &rects_optimal, vector <Mat>& output_area )

{

float r,angle;

for (int i = 0 ;i< rects_optimal.size() ; ++i)

{

//旋转区域

angle = rects_optimal[i].angle;

r = (float)rects_optimal[i].size.width / (float) (float)rects_optimal[i].size.height;

if(r<1)

angle = 90 + angle;

Mat rotmat = getRotationMatrix2D(rects_optimal[i].center , angle,1);//获得变形矩阵对象

Mat img_rotated;

warpAffine(intputImg ,img_rotated,rotmat, intputImg.size(),CV_INTER_CUBIC);

//裁剪图像

Size rect_size = rects_optimal[i].size;

if(r<1)

swap(rect_size.width, rect_size.height);

Mat img_crop;

getRectSubPix(img_rotated ,rect_size,rects_optimal[i].center , img_crop );

//用光照直方图调整所有裁剪得到的图像，使具有相同宽度和高度，适用于训练和分类

Mat resultResized;

resultResized.create(33,144,CV_8UC3);

resize(img_crop , resultResized,resultResized.size() , 0,0,INTER_CUBIC);

Mat grayResult;

RgbConvToGray(resultResized ,grayResult);

//blur(grayResult ,grayResult,Size(3,3));

equalizeHist(grayResult,grayResult);

output_area.push_back(grayResult);

}

}

void svm_train(CvSVM & svmClassifier)

{

FileStorage fs;

fs.open("SVM.xml" , FileStorage::READ);

Mat SVM_TrainningData;

Mat SVM_Classes;

fs["TrainingData"] >>SVM_TrainningData;

fs["classes"] >>SVM_Classes;

CvSVMParams SVM_params;

SVM_params.kernel_type = CvSVM::LINEAR;

svmClassifier.train(SVM_TrainningData,SVM_Classes ,Mat(),Mat(),SVM_params); //SVM训练模型

fs.release();

}

void char_segment(const Mat & inputImg,vector <Mat>& dst_mat)//得到20*20的标准字符分割图像

{

Mat img_threshold;

threshold(inputImg ,img_threshold , 180,255 ,CV_THRESH_BINARY );

Mat img_contours;

img_threshold.copyTo(img_contours);

vector < vector <Point> > contours;

findContours(img_contours ,contours,CV_RETR_EXTERNAL,CV_CHAIN_APPROX_NONE);

vector< vector <Point> > ::iterator itc = contours.begin();

vector<RotatedRect> char_rects;

while( itc != contours.end())

{

RotatedRect minArea = minAreaRect(Mat( *itc )); //返回每个轮廓的最小有界矩形区域

Point2f vertices[4];

minArea.points(vertices);

if(!char_verifySizes(minArea)) //判断矩形轮廓是否符合要求

{

itc = contours.erase(itc);

}

else

{

++itc;

char_rects.push_back(minArea);

}

}

char_sort(char_rects); //对字符排序

vector <Mat> char_mat;

for (int i = 0; i<char_rects.size() ;++i )

{

char_mat.push_back( Mat(img_threshold,char_rects[i].boundingRect())) ;

}

Mat train_mat(2,3,CV_32FC1);

int length ;

dst_mat.resize(7);

Point2f srcTri[3];

Point2f dstTri[3];

for (int i = 0; i< char_mat.size();++i)

{

srcTri[0] = Point2f( 0,0 );

srcTri[1] = Point2f( char_mat[i].cols - 1, 0 );

srcTri[2] = Point2f( 0, char_mat[i].rows - 1 );

length = char_mat[i].rows > char_mat[i].cols?char_mat[i].rows:char_mat[i].cols;

dstTri[0] = Point2f( 0.0, 0.0 );

dstTri[1] = Point2f( length, 0.0 );

dstTri[2] = Point2f( 0.0, length );

train_mat = getAffineTransform( srcTri, dstTri );

dst_mat[i]=Mat::zeros(length,length,char_mat[i].type());

warpAffine(char_mat[i],dst_mat[i],train_mat,dst_mat[i].size(),INTER_LINEAR,BORDER_CONSTANT,Scalar(0));

resize(dst_mat[i],dst_mat[i],Size(20,20)); //尺寸调整为20*20

}

}

bool char_verifySizes(const RotatedRect & candidate)

{

float aspect = 33.0f/20.0f;

float charAspect = (float) candidate.size.width/ (float)candidate.size.height; //宽高比

float error = 0.35;

float minHeight = 11; //最小高度11

float maxHeight = 33; //最大高度33

float minAspect = 0.20; //考虑到数字1，最小长宽比为0.15

float maxAspect = aspect + aspect*error;

if( charAspect > minAspect && charAspect < maxAspect

&& candidate.size.height >= minHeight && candidate.size.width< maxHeight) //非0像素点数、长宽比、高度需满足条件

return true;

else

return false;

}

void char_sort(vector <RotatedRect > & in_char ) //对字符区域进行排序

{

vector <RotatedRect > out_char;

const int length = 7; //7个字符

int index[length] = {0,1,2,3,4,5,6};

float centerX[length];

for (int i=0;i < length ; ++ i)

{

centerX[i] = in_char[i].center.x;

}

for (int j=0;j <length;j++) {

for (int i=length-2;i >= j;i--)

if (centerX[i] > centerX[i+1])

{

float t=centerX[i];

centerX[i]=centerX[i+1];

centerX[i+1]=t;

int tt = index[i];

index[i] = index[i+1];

index[i+1] = tt;

}

}

for(int i=0;i<length ;i++)

out_char.push_back(in_char[(index[i])]);

in_char.clear(); //清空in_char

in_char = out_char; //将排序好的字符区域向量重新赋值给in_char

}

void features(const Mat & in , Mat & out ,int sizeData)

{

Mat vhist = projectHistogram(in , 1); //水平直方图

Mat hhist = projectHistogram(in , 0); //垂直直方图

Mat lowData;

resize(in , lowData ,Size(sizeData ,sizeData ));

int numCols = vhist.cols + hhist.cols + lowData.cols * lowData.cols;

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);

}

for(int x =0 ;x<lowData.rows ;++x)

{

for (int y =0 ;y < lowData.cols ;++ y)

{

out.at<float>(j) = (float)lowData.at<unsigned char>(x,y);

j++;

}

}

}

Mat projectHistogram(const Mat& img ,int t) //水平或垂直直方图,0为按列统计

{ //1为按行统计

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);

}

double min,max;

minMaxLoc(mhist , &min ,&max);

if(max > 0)

mhist.convertTo(mhist ,-1,1.0f/max , 0);

return mhist;

}

void ann_train(CvANN_MLP &ann ,int numCharacters, int nlayers)

{

Mat trainData ,classes;

FileStorage fs;

fs.open("ann_xml.xml" , FileStorage::READ);

fs["TrainingData"] >>trainData;

fs["classes"] >>classes;

Mat layerSizes(1,3,CV_32SC1);

layerSizes.at<int>( 0 ) = trainData.cols;

layerSizes.at<int>( 1 ) = nlayers; //隐藏神经元数，可设为3

layerSizes.at<int>( 2 ) = numCharacters; //样本类数为34

ann.create(layerSizes , CvANN_MLP::SIGMOID_SYM ,1,1); //初始化ann

Mat trainClasses;

trainClasses.create(trainData.rows , numCharacters ,CV_32FC1);

for (int i =0;i< trainData.rows; i++)

{

for (int k=0 ; k< trainClasses.cols ; k++ )

{

if ( k == (int)classes.at<uchar> (i))

{

trainClasses.at<float>(i,k) = 1 ;

}

else

trainClasses.at<float>(i,k) = 0;

}

}

Mat weights(1 , trainData.rows , CV_32FC1 ,Scalar::all(1) );

ann.train( trainData ,trainClasses , weights);

}

void classify(CvANN_MLP& ann, vector<Mat> &char_feature , vector<int> & char_result)

{

char_result.resize(char_feature.size());

for (int i=0;i<char_feature.size(); ++i)

{

Mat output(1 ,34 , CV_32FC1); //1*34矩阵

ann.predict(char_feature[i] ,output);

Point maxLoc;

double maxVal;

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

char_result[i] = maxLoc.x;

}

std::cout<<"该车牌后6位为：";

char s[] = {'0','1','2','3','4','5','6','7','8','9','A','B',

'C','D','E','F','G','H','J','K','L','M','N','P','Q',

'R','S','T','U','V','W','X','Y','Z'};

for (int i=1;i<char_result.size(); ++i) //第一位是汉字，这里没实现对汉字的预测

{

std::cout<<s[char_result[i]];

}

std::cout<<'\n';

}

四、关于SVM.xml与ann_xml.xml

SVM.xml中保存的是用于SVM训练的训练矩阵和类别矩阵数据，标签为"TrainingData"对应训练矩阵，为195*4752大小，195表示有195个训练样本，4752表示每个样本的特征向量维度，因为每个图片大小为144*33，将其转为一行，也即将每个像素值都作为一个特征值，则可得到4752个特征值，保存为1行，所使用的样本图片有75个正样本和120个负样本，保存如下：



标签为"classes"对应类别矩阵，为195*1矩阵，前75个值对应正样本为1.0，后120个值对应负样本为-1.0。

同理，ann_xml.xml保存的是用于神经网络训练的数据，标签为"TrainingData"对应训练矩阵，为1700*65大小，1700表示训练样本数，因为车牌字符有34类，每类有50个，故总数1700，65为每个样本提取的特征向量。标签为"classes"对应类别矩阵，为1700*1大小，标记负样本数是哪一类的，训练样本取自



五、结果及分析

勉强找到2张图片，可以完整地识别出车牌的后6为字符，效果如下。





故该系统的性能仍有待提升，不过我认为在车牌定位方面可以有所改进，采用其他更好的车牌定位算法会更好，以及可以增加神经网络的训练样本。

六、完整代码文件下载

下载链接为：http://download.csdn.NET/detail/ap1005834/9513328