暗通道算法是由何恺明在CVPR论文《Single ImageHaze Removalusing Dark Channel Prior》中提出的。
图像去雾的模型:
我们分析以上模型: 【已知条件】 :observerd intensity,即输入图像(待去雾的图像) 【未知条件】 scene radiance,即还原图像(去雾以后的图像) medium transmission global atmospheric light 【目标】 求出这三个未知条件、、,而根据去雾模型,我们只需要计算出其中两个未知条件,就可以求出第三个。文中先通过求出、,然后通过去雾模型的转换计算 【问题】 怎么只通过,来计算出和呢? 【问题的解决办法】 也就是我们先求出darkChannel. darkChannel的定义: 代码如下: [cpp] view plain copy print ? #include<iostream> #include<vector> #include<algorithm> using namespace std; #include<opencv2\core\core.hpp> #include<opencv2\highgui\highgui.hpp> #include<opencv2\imgproc\imgproc.hpp> using namespace cv; int main(int argc,char*argv[]) { Mat image=imread(argv[1],1); CV_Assert(!image.empty() && image.channels() == 3); //图片的归一化 Mat fImage; image.convertTo(fImage,CV_32FC3,1.0/255,0); //规定patch的大小,且均为奇数 int hPatch = 15; int vPatch = 15; //给归一化的图片添加边界 Mat fImageBorder; copyMakeBorder(fImage,fImageBorder,vPatch/2,vPatch/2,hPatch/2,hPatch/2,BORDER_REPLICATE); //分离通道 vector<Mat> fImageBorderVector(3); split(fImageBorder,fImageBorderVector); //创建darkChannel Mat darkChannel(image.rows,image.cols,CV_32FC1); double minTemp ,minPixel; //根据darkChannel的定义 for(unsigned int r = 0;r < darkChannel.rows;r++) { for(unsigned int c = 0;c < darkChannel.cols;c++) { minPixel = 1.0; for(vector<Mat>::iterator it = fImageBorderVector.begin() ;it != fImageBorderVector.end();it++) { Mat roi(*it,Rect(c,r,hPatch,vPatch)); minMaxLoc(roi,&minTemp); minPixel = min(minPixel,minTemp); } darkChannel.at<float>(r,c) = float(minPixel); } } namedWindow("darkChannel",1); imshow("darkChannel",darkChannel); Mat darkChannel8U; darkChannel.convertTo(darkChannel8U,CV_8UC1,255,0); imwrite("darkChannel.jpg",darkChannel8U); return 0; } 先给出一些运行结果: 第二步:通过暗通道来实现A的过程, [cpp] view plain copy print ? /*第2步:求出 A(global atmospheric light)*/ //2.1 计算出darkChannel中,前top个亮的值,论文中取值为0.1% float top = 0.001; float numberTop = top*darkChannel.rows*darkChannel.cols; Mat darkChannelVector; darkChannelVector = darkChannel.reshape(1,1); Mat_<int> darkChannelVectorIndex; sortIdx(darkChannelVector,darkChannelVectorIndex,CV_SORT_EVERY_ROW + CV_SORT_DESCENDING); //制作掩码 Mat mask(darkChannelVectorIndex.rows,darkChannelVectorIndex.cols,CV_8UC1);//注意mask的类型必须是CV_8UC1 for(unsigned int r = 0;r < darkChannelVectorIndex.rows;r++) { for(unsigned int c = 0;c < darkChannelVectorIndex.cols;c++) { if(darkChannelVectorIndex.at<int>(r,c) <= numberTop) mask.at<uchar>(r,c) = 1; else mask.at<uchar>(r,c) = 0; } } Mat darkChannelIndex = mask.reshape(1,darkChannel.rows); vector<double> A(3);//分别存取A_b,A_g,A_r vector<double>::iterator itA = A.begin(); vector<Mat>::iterator it = fImageBorderVector.begin(); //2.2在求第三步的t(x)时,会用到以下的矩阵,这里可以提前求出 vector<Mat> fImageBorderVectorA(3); vector<Mat>::iterator itAA = fImageBorderVectorA.begin(); for( ;it != fImageBorderVector.end() && itA != A.end() && itAA != fImageBorderVectorA.end();it++,itA++,itAA++) { Mat roi(*it,Rect(hPatch/2,vPatch/2,darkChannel.cols,darkChannel.rows)); minMaxLoc(roi,0,&(*itA),0,0,darkChannelIndex);// (*itAA) = (*it)/(*itA); //[注意:这个地方有除号,但是没有判断是否等于0] } 第三步:通过暗通道来实现t(x)的过程: [cpp] view plain copy print ? /*第三步:求t(x)*/ Mat darkChannelA(darkChannel.rows,darkChannel.cols,CV_32FC1); float omega = 0.95;//0<w<=1,论文中取值为0.95 //代码和求darkChannel的时候,代码差不多 for(unsigned int r = 0;r < darkChannel.rows;r++) { for(unsigned int c = 0;c < darkChannel.cols;c++) { minPixel = 1.0; for(itAA = fImageBorderVectorA.begin() ;itAA != fImageBorderVectorA.end();itAA++) { Mat roi(*itAA,Rect(c,r,hPatch,vPatch)); minMaxLoc(roi,&minTemp); minPixel = min(minPixel,minTemp); } darkChannelA.at<float>(r,c) = float(minPixel); } } Mat tx = 1.0 - omega*darkChannelA; 文中,给出了一个tx的优化,我们后面使用guiderFilter进行优化。 第四步:既然A和t(x)已经求出,就可以求j(x); [cpp] view plain copy print ? /*第四步:我们可以求J(x)*/ float t0 = 0.1;//论文中取t0 = 0.1 Mat jx(image.rows,image.cols,CV_32FC3); for(size_t r = 0;r < jx.rows;r++) { for(size_t c =0;c<jx.cols;c++) { jx.at<Vec3f>(r,c) = Vec3f((fImage.at<Vec3f>(r,c)[0] - A[0])/max(tx.at<float>(r,c),t0)+A[0],(fImage.at<Vec3f>(r,c)[1] - A[1])/max(tx.at<float>(r,c),t0)+A[1],(fImage.at<Vec3f>(r,c)[2] - A[2])/max(tx.at<float>(r,c),t0)+A[2]); } } namedWindow("jx",1); imshow("jx",jx); Mat jx8U; jx.convertTo(jx8U,CV_8UC3,255,0); imwrite("jx.jpg",jx8U); 结果: 文中的代码还没有优化,代码重复率比较高
转自:http://blog.csdn.net/zhangping1987/article/details/51178103