Opencv中meanShiftSegmentation的实现
Opencv中meanShiftSegmentation的实现
1.样例在opencv-2.4.6.1\samples\cpp的meanShift_Segmentation.cpp中
static void meanShiftSegmentation( int, void* ){
cout << "spatialRad=" << spatialRad << "; "
<< "colorRad=" << colorRad << "; "
<< "maxPyrLevel=" << maxPyrLevel << endl;
pyrMeanShiftFiltering( img, res, spatialRad, colorRad, maxPyrLevel );
floodFillPostprocess( res, Scalar::all(2) );
imshow( winName, res );
}
2. 下面着重介绍一下pyrMeanShiftFiltering.核心函数式cvPyrMeanShiftFiltering。
该函数使用金字塔的方法加速，每一级金字塔中实际依然是使用meanshift。核心思想如下：
(1) 从金子塔塔顶的图像开始处理：（塔顶的图像最小，速度较快）。图像的每个点进行以下meanshift迭代运算：
a) 从图像左上角第一个像素开始，以该点为中心，生成指定大小的窗口。计算窗口中满足距离条件的所有点的平均位置以及R,G,B的平均值。
（距离条件：（R1-R0）*（R1-R0）+(G1-G0)*(G1-G0)+(B1-B0)*(B1-B0) 其中R0，G0，B0为窗口中心像素的RGB值;R1,G1,B1为窗口中每个像素相应的RGB值） b) 将窗口中心平移到a)获取的平均位置，计算新窗口中满足距离条件的所有点的新平均位置以及新R,G,B的平均值。（R0，G0，B0为a)中获取的新R,G,B的平均值，即距离比较的参考点为a)中获取的新R,G,B的平均值。） c) 不断重复执行b),直到满足迭代次数或者新平均位置和新R,G,B的平均值和上一次的值差距满足迭代精度。另外在重复执行b)时，每一次都将窗口中心不断平移到b)获取的新平均位置，而且距离条件中比较的参考点为b)获取的新R,G,B的平均值。 d) 将迭代结束获得的新R,G,B的平均值存储到窗口中心位置对应的像素中。 不断移动窗口中心，将图像的每个点都做为窗口中心进行遍历,执行上面的a,b，c，d。 (2) 将（1）的结果（新的金字塔图像）生成下一层金字塔计算的初值和mask。利用mask实现加速，只有mask为1的位置才进行本轮的迭代。 a) 将（1）的结果升2采样，即图像长宽都扩大2倍，整个图像扩大4倍。将其作为改层金字塔图像计算的初值。 b) 判断a)结果中的所有像素点是否满足差异条件，如果满足则mask为1，否则mask为0。 （判断条件：（R1-R0）*（R1-R0）+(G1-G0)*(G1-G0)+(B1-B0)*(B1-B0) 其中R0，G0，B0为窗口中心像素的RGB值;R1,G1,B1为窗口中每个像素相应的RGB值。。。。。。。。） 待续。。。。。 (3) 根据（2）中的初值和mask，将mask为1的所有像素进行meanshift迭代，迭代过程和步骤（1）一样，最后把迭代结束获得的新R,G,B的平均值存储到mask为1的位置的对应像素中。 (4) 重复步骤（2）和步骤（3），直至金字塔底。最终输出和原图大小一样的结果图像。 CV_IMPL void cvPyrMeanShiftFiltering( const CvArr* srcarr, CvArr* dstarr, double sp0, double sr, int max_level, CvTermCriteria termcrit ) { const int cn = 3; const int MAX_LEVELS = 8; if( (unsigned)max_level > (unsigned)MAX_LEVELS ) CV_Error( CV_StsOutOfRange, "The number of pyramid levels is too large or negative" ); std::vector<cv::Mat> src_pyramid(max_level+1); std::vector<cv::Mat> dst_pyramid(max_level+1); cv::Mat mask0; int i, j, level; //uchar* submask = 0; #define cdiff(ofs0) (tab[c0-dptr[ofs0]+255] + \ tab[c1-dptr[(ofs0)+1]+255] + tab[c2-dptr[(ofs0)+2]+255] >= isr22) double sr2 = sr * sr; int isr2 = cvRound(sr2), isr22 = MAX(isr2,16); int tab[768]; cv::Mat src0 = cv::cvarrToMat(srcarr); cv::Mat dst0 = cv::cvarrToMat(dstarr); if( src0.type() != CV_8UC3 ) CV_Error( CV_StsUnsupportedFormat, "Only 8-bit, 3-channel images are supported" ); if( src0.type() != dst0.type() ) CV_Error( CV_StsUnmatchedFormats, "The input and output images must have the same type" ); if( src0.size() != dst0.size() ) CV_Error( CV_StsUnmatchedSizes, "The input and output images must have the same size" ); if( !(termcrit.type & CV_TERMCRIT_ITER) ) termcrit.max_iter = 5; termcrit.max_iter = MAX(termcrit.max_iter,1); termcrit.max_iter = MIN(termcrit.max_iter,100); if( !(termcrit.type & CV_TERMCRIT_EPS) ) termcrit.epsilon = 1.f; termcrit.epsilon = MAX(termcrit.epsilon, 0.f); for( i = 0; i < 768; i++ ) tab[i] = (i - 255)*(i - 255); // 1. construct pyramid src_pyramid[0] = src0; dst_pyramid[0] = dst0; for( level = 1; level <= max_level; level++ ) { src_pyramid[level].create( (src_pyramid[level-1].rows+1)/2, (src_pyramid[level-1].cols+1)/2, src_pyramid[level-1].type() ); dst_pyramid[level].create( src_pyramid[level].rows, src_pyramid[level].cols, src_pyramid[level].type() ); cv::pyrDown( src_pyramid[level-1], src_pyramid[level], src_pyramid[level].size() ); //CV_CALL( cvResize( src_pyramid[level-1], src_pyramid[level], CV_INTER_AREA )); } mask0.create(src0.rows, src0.cols, CV_8UC1); //CV_CALL( submask = (uchar*)cvAlloc( (sp+2)*(sp+2) )); // 2. apply meanshift, starting from the pyramid top (i.e. the smallest layer) for( level = max_level; level >= 0; level-- ) { cv::Mat src = src_pyramid[level]; cv::Size size = src.size(); uchar* sptr = src.data; int sstep = (int)src.step; uchar* mask = 0; int mstep = 0; uchar* dptr; int dstep; float sp = (float)(sp0 / (1 << level)); sp = MAX( sp, 1 ); if( level < max_level ) { cv::Size size1 = dst_pyramid[level+1].size(); cv::Mat m( size.height, size.width, CV_8UC1, mask0.data ); dstep = (int)dst_pyramid[level+1].step; dptr = dst_pyramid[level+1].data + dstep + cn; mstep = (int)m.step; mask = m.data + mstep; //cvResize( dst_pyramid[level+1], dst_pyramid[level], CV_INTER_CUBIC ); cv::pyrUp( dst_pyramid[level+1], dst_pyramid[level], dst_pyramid[level].size() ); m.setTo(cv::Scalar::all(0)); for( i = 1; i < size1.height-1; i++, dptr += dstep - (size1.width-2)*3, mask += mstep*2 ) { for( j = 1; j < size1.width-1; j++, dptr += cn ) { int c0 = dptr[0], c1 = dptr[1], c2 = dptr[2]; mask[j*2 - 1] = cdiff(-3) || cdiff(3) || cdiff(-dstep-3) || cdiff(-dstep) || cdiff(-dstep+3) || cdiff(dstep-3) || cdiff(dstep) || cdiff(dstep+3); } } cv::dilate( m, m, cv::Mat() ); mask = m.data; } dptr = dst_pyramid[level].data; dstep = (int)dst_pyramid[level].step; for( i = 0; i < size.height; i++, sptr += sstep - size.width*3, dptr += dstep - size.width*3, mask += mstep ) { for( j = 0; j < size.width; j++, sptr += 3, dptr += 3 ) { int x0 = j, y0 = i, x1, y1, iter; int c0, c1, c2; if( mask && !mask[j] ) continue; c0 = sptr[0], c1 = sptr[1], c2 = sptr[2]; // iterate meanshift procedure，图像的每点都使用meanshift，找到其收敛的RGB for( iter = 0; iter < termcrit.max_iter; iter++ ) { uchar* ptr; int x, y, count = 0; int minx, miny, maxx, maxy; int s0 = 0, s1 = 0, s2 = 0, sx = 0, sy = 0; double icount; int stop_flag; //mean shift: process pixels in window (p-sigmaSp)x(p+sigmaSp) minx = cvRound(x0 - sp); minx = MAX(minx, 0); miny = cvRound(y0 - sp); miny = MAX(miny, 0); maxx = cvRound(x0 + sp); maxx = MIN(maxx, size.width-1); maxy = cvRound(y0 + sp); maxy = MIN(maxy, size.height-1); ptr = sptr + (miny - i)*sstep + (minx - j)*3; for( y = miny; y <= maxy; y++, ptr += sstep - (maxx-minx+1)*3 ) { int row_count = 0; x = minx; #if CV_ENABLE_UNROLLED for( ; x + 3 <= maxx; x += 4, ptr += 12 ) { int t0 = ptr[0], t1 = ptr[1], t2 = ptr[2]; if( tab[t0-c0+255] + tab[t1-c1+255] + tab[t2-c2+255] <= isr2 ) { s0 += t0; s1 += t1; s2 += t2; sx += x; row_count++; } t0 = ptr[3], t1 = ptr[4], t2 = ptr[5]; if( tab[t0-c0+255] + tab[t1-c1+255] + tab[t2-c2+255] <= isr2 ) { s0 += t0; s1 += t1; s2 += t2; sx += x+1; row_count++; } t0 = ptr[6], t1 = ptr[7], t2 = ptr[8]; if( tab[t0-c0+255] + tab[t1-c1+255] + tab[t2-c2+255] <= isr2 ) { s0 += t0; s1 += t1; s2 += t2; sx += x+2; row_count++; } t0 = ptr[9], t1 = ptr[10], t2 = ptr[11]; if( tab[t0-c0+255] + tab[t1-c1+255] + tab[t2-c2+255] <= isr2 ) { s0 += t0; s1 += t1; s2 += t2; sx += x+3; row_count++; } } #endif for( ; x <= maxx; x++, ptr += 3 ) { int t0 = ptr[0], t1 = ptr[1], t2 = ptr[2]; if( tab[t0-c0+255] + tab[t1-c1+255] + tab[t2-c2+255] <= isr2 ) { s0 += t0; s1 += t1; s2 += t2; sx += x; row_count++; } } count += row_count; sy += y*row_count; } if( count == 0 ) break; icount = 1./count; x1 = cvRound(sx*icount); y1 = cvRound(sy*icount); s0 = cvRound(s0*icount); s1 = cvRound(s1*icount); s2 = cvRound(s2*icount); stop_flag = (x0 == x1 && y0 == y1) || abs(x1-x0) + abs(y1-y0) + tab[s0 - c0 + 255] + tab[s1 - c1 + 255] + tab[s2 - c2 + 255] <= termcrit.epsilon; x0 = x1; y0 = y1; c0 = s0; c1 = s1; c2 = s2; if( stop_flag ) break; } dptr[0] = (uchar)c0; dptr[1] = (uchar)c1; dptr[2] = (uchar)c2; } }// 一层 } }