学步园

此验证码识别技术有预处理、特征提取、神经网络识别几部分，参考过车牌识别技术。程序实现分样本下载器、用户交互程序、MLP建构程序三块，使用OpenCV，喜欢的可以留意下AnnMLP的使用方法，国内这方面的资料还是比较少的，欢迎交流QQ:13895354。

预处理技术是专门针对某种特定验证码设计的，其特点有：英文+数字，彩色，倾斜，无粘连，有不规则噪声，有删除线，至于具体哪家网站的验证码，我就不方便透露了

（用户交互程序）

#include <cv.h><br /> #include <highgui.h><br /> #include <ml.h><br /> #include <iostream><br /> using namespace std;</p> <p>int main( int argc, char** argv )<br /> {<br /> cvNamedWindow("1");<br /> cvNamedWindow("2");<br /> IplConvKernel* se = cvCreateStructuringElementEx(2, 2, 1, 1, CV_SHAPE_CROSS);</p> <p> CvANN_MLP mlp;<br /> mlp.load( "mpl.xml" );</p> <p> for(int i=1; i<=100; i++)<br /> {<br /> // 产生文件名<br /> char fn[1024];<br /> sprintf(fn, ".//sample//%d.jpg", i);</p> <p> // 读取图片<br /> IplImage* gray = cvLoadImage(fn, CV_LOAD_IMAGE_GRAYSCALE);</p> <p> // 二值化<br /> cvThreshold(gray, gray, 180, 255, CV_THRESH_BINARY | CV_THRESH_OTSU);<br /> // 去边框<br /> cvRectangle(gray, cvPoint(0, 0), cvPoint(gray->width-1, gray->height-1), CV_RGB(255, 255, 255));<br /> // 去噪<br /> cvDilate(gray, gray, se);</p> <p> /*<br /> // 调整角度<br /> cvShowImage("1", gray);<br /> IplImage* rote = cvCreateImage( cvGetSize(gray), IPL_DEPTH_8U, 1 );<br /> double t = tan(10.0 / 180.0 * CV_PI);<br /> int w = gray->width;<br /> int h = gray->height;<br /> for(int i = 0; i<h; i++)<br /> {<br /> unsigned char* lineGray = (unsigned char*)gray->imageData + gray->widthStep * i;<br /> unsigned char* lineRote = (unsigned char*)rote->imageData + rote->widthStep * i;<br /> for(int j = 0; j<w; j++)<br /> {<br /> int j2 = j - ((int)(i*t+0.5));<br /> if (j2<0)<br /> j2+=w;<br /> *(lineRote+j) = *(lineGray+j2);<br /> }<br /> }<br /> cvCopy(rote, gray);<br /> cvReleaseImage(&rote);<br /> */</p> <p> // 计算连通域contour<br /> cvXorS(gray, cvScalarAll(255), gray, 0);<br /> cvShowImage("1", gray);<br /> CvMemStorage* storage = cvCreateMemStorage();<br /> CvSeq* contour = NULL;<br /> cvFindContours(gray, storage, &contour, sizeof(CvContour), CV_RETR_CCOMP, CV_CHAIN_APPROX_SIMPLE);</p> <p> // 分析连通域<br /> CvSeq* p = contour;<br /> while(p)<br /> {<br /> CvRect rect = cvBoundingRect(p, 0);<br /> if( rect.height>=10 )// 文字需要有10像素高度<br /> {<br /> // 绘制该连通区域到character<br /> cvZero(gray);<br /> IplImage* character = cvCreateImage(cvSize(rect.width, rect.height), IPL_DEPTH_8U, 1);<br /> cvZero(character);<br /> cvDrawContours(character, p, CV_RGB(255, 255, 255), CV_RGB(0, 0, 0), -1, -1, 8, cvPoint(-rect.x, -rect.y));</p> <p> // 归一化<br /> IplImage* normal = cvCreateImage(cvSize(16, 16), IPL_DEPTH_8U, 1);<br /> cvResize(character, normal, CV_INTER_AREA);<br /> cvThreshold(normal, normal, 0, 255, CV_THRESH_BINARY | CV_THRESH_OTSU);// 修正</p> <p> // 计算输入向量<br /> float input[256];<br /> for(int i=0; i<256; i++)<br /> input[i] = (normal->imageData[i]==-1);</p> <p>#if 0<br /> // 用户输入结果<br /> cvShowImage("2", normal);<br /> char c = cvWaitKey(0);<br /> if(c==27)<br /> return 0;</p> <p> // 编码0-9:0-9 a-z:10-35<br /> unsigned char cc = 255;<br /> if(c>='A'&&c<='Z')<br /> cc=c-'A'+10;<br /> else if(c>='a'&&c<='z')<br /> cc=c-'a'+10;<br /> else if(c>='0'&&c<='9')<br /> cc=c-'0';</p> <p> if(cc!=255)<br /> {<br /> // 转换成输出向量<br /> float output[36];<br /> for(int i=0; i<36; i++)<br /> output[i] = 0.0f;<br /> output[cc] = 1.0f;</p> <p> // 存储到批处理文件<br /> static FILE* batch = fopen("batch", "wb");<br /> fwrite(output, 4*36, 1, batch);<br /> fwrite(input, 4*256, 1, batch);</p> <p> static int count = 0;<br /> cout<<count++<<endl;<br /> }</p> <p>#else<br /> // 识别<br /> CvMat* output = cvCreateMat( 1, 36, CV_32F );<br /> CvMat inputMat = cvMat( 1, 256, CV_32F, input);<br /> mlp.predict( &inputMat, output );<br /> CvPoint max_loc = {0,0};<br /> cvMinMaxLoc( output, NULL, NULL, NULL, &max_loc, NULL );<br /> int best = max_loc.x;// 识别结果<br /> char c = (char)( best<10 ? '0'+best : 'A'+best-10 );<br /> cout<<c<<"("<<rect.x<<")"<<" ";<br /> cvReleaseMat( &output );</p> <p>#endif<br /> cvReleaseImage(&character);<br /> cvReleaseImage(&normal);<br /> }<br /> p = p->h_next;<br /> }<br /> cout<<endl;<br /> cvWaitKey(0);<br /> cvReleaseMemStorage(&storage);<br /> cvReleaseImage(&gray);<br /> }</p> <p> cvReleaseStructuringElement(&se);<br /> cvDestroyAllWindows();</p> <p> return 0;<br /> }<br />

事实上更有价值的是OpenCV中的AnnMLP，其技术很成熟，参考了letter_recg范例程序

 （MLP建构程序）

#include <iostream><br /> #include <ml.h><br /> using namespace std;</p> <p>void print_mat(CvMat& mat)<br /> {<br /> int count = 0;<br /> for(int i=0; i<4/*mat.rows*/; i++)<br /> {<br /> for(int j=0; j<mat.cols; j++)<br /> {<br /> cout<<mat.data.fl[i*(mat.step/4)+j]/*<<" "*/;<br /> }<br /> cout<<endl<<endl;<br /> }<br /> }</p> <p>int main( int argc, char *argv[] )<br /> {<br /> // 读入结果responses 特征data<br /> FILE* f = fopen( "batch", "rb" );<br /> fseek(f, 0l, SEEK_END);<br /> long size = ftell(f);<br /> fseek(f, 0l, SEEK_SET);<br /> int count = size/4/(36+256);<br /> CvMat* batch = cvCreateMat( count, 36+256, CV_32F );<br /> fread(batch->data.fl, size-1, 1, f);<br /> CvMat outputs, inputs;<br /> cvGetCols(batch, &outputs, 0, 36);<br /> cvGetCols(batch, &inputs, 36, 36+256);</p> <p> // 新建MPL<br /> CvANN_MLP mlp;<br /> int layer_sz[] = { 256, 20, 36 };<br /> CvMat layer_sizes = cvMat( 1, 3, CV_32S, layer_sz );<br /> mlp.create( &layer_sizes );</p> <p> // 训练<br /> system( "time" );<br /> mlp.train( &inputs, &outputs, NULL, NULL,<br /> CvANN_MLP_TrainParams(cvTermCriteria(CV_TERMCRIT_ITER,300,0.01), CvANN_MLP_TrainParams::RPROP, 0.01)<br /> );<br /> system( "time" );</p> <p> // 存储MPL<br /> mlp.save( "mpl.xml" );</p> <p> // 测试<br /> int right = 0;<br /> CvMat* output = cvCreateMat( 1, 36, CV_32F );<br /> for(int i=0; i<count; i++)<br /> {<br /> CvMat input;<br /> cvGetRow( &inputs, &input, i );</p> <p> mlp.predict( &input, output );<br /> CvPoint max_loc = {0,0};<br /> cvMinMaxLoc( output, NULL, NULL, NULL, &max_loc, NULL );<br /> int best = max_loc.x;// 识别结果</p> <p> int ans = -1;// 实际结果<br /> for(int j=0; j<36; j++)<br /> {<br /> if( outputs.data.fl[i*(outputs.step/4)+j] == 1.0f )<br /> {<br /> ans = j;<br /> break;<br /> }<br /> }<br /> cout<<(char)( best<10 ? '0'+best : 'A'+best-10 );<br /> cout<<(char)( ans<10 ? '0'+ans : 'A'+ans-10 );<br /> if( best==ans )<br /> {<br /> cout<<"+";<br /> right++;<br /> }<br /> //cin.get();<br /> cout<<endl;<br /> }<br /> cvReleaseMat( &output );<br /> cout<<endl<<right<<"/"<<count<<endl;</p> <p> cvReleaseMat( &batch );</p> <p> system( "pause" );</p> <p> return 0;<br /> }<br />

上面的代码都是vc6控制台程序，OpenCV入门的都懂的吧。