一、准备样本:
正样本:样本处理和选择非常有技巧,我的标准是框住整个车牌留出边框,这样既保留了车牌原有的字符特征,字符组特征还有车牌的边框特征。其中双行车牌我只取底下面的一行。并且检测样本最好不要预处理,输入源给出什么图形就用什么图形。具体的抠图方式可以参考我其他博客车牌识别技术详解三--字符检测的正负样本得取(利用鼠标画框抠图)。
负样本:负样本选择同样非常有技巧性。尽量采集车牌使用环境下的背景图片,并且需要包含一部分车牌字符但是非正样本的取在车牌周围的负样本。
二、HAAR特征:
可视化haar特征
haarFeatureVision
改动:
(1)global.h里面
extern IplImage * showHaarFeatureImg;
(2)global.cpp里面
IplImage * showHaarFeatureImg;
(3)SimpleClassifier.cpp里面改动
#define haarFeatureVison 1
void SimpleClassifier::ReadFromFile(ifstream& f)
{
f>>thresh>>parity>>type;
f>>x1>>x2>>x3>>x4>>y1>>y2>>y3>>y4;
f.ignore(256,'\n');
assert(parity == 0 || parity == 1);
// assert(type>=0 && type<=4);
if(haarFeatureVison){//haarFeatureVison
showHaarFeatureImg = cvLoadImage("showHaarFeature.bmp",1);
if(type == 0){
std::cout<<x1<<" "<<x2<<" "<<x3<<" "<<x4<<" "<<y1<<" "<<y2<<" "<<y3<<" "<<y4<<std::endl;
CvPoint A = cvPoint(y1,x1);
CvPoint B = cvPoint(y1,x2);
CvPoint C = cvPoint(y1,x3);
CvPoint D = cvPoint(y3,x1);
CvPoint E = cvPoint(y3,x2);
CvPoint F = cvPoint(y3,x3);
cvRectangle(showHaarFeatureImg, A,cvPoint(E.x,E.y-1),CV_RGB(0,0,0), CV_FILLED);
cvRectangle(showHaarFeatureImg, B,cvPoint(F.x,F.y-1),CV_RGB(200,200,200), CV_FILLED);
cvNamedWindow("haarFeatureVison",0);
cvShowImage("haarFeatureVison",showHaarFeatureImg);
// cvWaitKey(0);
}
if(type == 5){
std::cout<<x1<<" "<<x2<<" "<<x3<<" "<<x4<<" "<<y1<<" "<<y2<<" "<<y3<<" "<<y4<<std::endl;
CvPoint A = cvPoint(y1,x1);
CvPoint B = cvPoint(y4,x1);
CvPoint C = cvPoint(y1,x2);
CvPoint D = cvPoint(y2,x2);
CvPoint E = cvPoint(y3,x2);
CvPoint F = cvPoint(y4,x2);
CvPoint G = cvPoint(y1,x3);
CvPoint H = cvPoint(y2,x3);
CvPoint I = cvPoint(y3,x3);
CvPoint J = cvPoint(y4,x3);
CvPoint K = cvPoint(y1,x4);
CvPoint L = cvPoint(y4,x4);
cvRectangle(showHaarFeatureImg, D,cvPoint(I.x-1, I.y-1),CV_RG