立體匹配中的NCC，SAD，SSD算法

http://www.cnblogs.com/cvlabs/archive/2010/01/13/1646902.html

Normalized Cross correlation (NCC)

NCC(u,v) =  [(wl - w)/(|wl - w|)]*[(wr - w)/(|wr - w|)] 選擇最大值

 

Sum of Squared Defferences (SSD)

SSD(u,v) =  Sum{[Left(u,v) - Right(u,v)] * [Left(u,v) - Right(u,v)]} 選擇最大值

Sum of Absolute Defferences (SAD) 

SAD(u,v) = Sum{|Left(u,v) - Right(u,v)|}  選擇最小值

 

先說說SAD算法的基本流程：

1.構造一個小窗口，類似與卷積核。

2.用窗口覆蓋左邊的圖像，選擇出窗口覆蓋區域內的所有像素點。 

3.同樣用窗口覆蓋右邊的圖像並選擇出覆蓋區域的像素點。

4.左邊覆蓋區域減去右邊覆蓋區域，並求出所有像素點差的絕對值的和。

5.移動右邊圖像的窗口，重復3，4的動作。（這裡有個搜索范圍，超過這個范圍跳出） 

6.找到這個范圍內SAD值最小的窗口，即找到了左邊圖像的最佳匹配的像素塊。

 

OpenCV代碼示范SAD:

代碼Code highlighting produced by Actipro CodeHighlighter (freeware)http://www.CodeHighlighter.com/--> 1 IplImage* generateDisparityImage(IplImage* greyLeftImg32,
                          IplImage* greyRightImg32, 
                          int windowSize,int DSR){
 
     int offset=floor((double)windowSize/2); 
    int height=greyLeftImg32->height;
    int width=greyLeftImg32->width;
     double* localSAD=new double[DSR];//DSR即搜索范圍
 
    int x=0, y=0,d=0,m=0;
    int N=windowSize;             
 
    IplImage* winImg=cvCreateImage(cvSize(N,N),32,1);//mySubImage(greyLeftImg32,cvRect(0,0,N,N));
    
    IplImage* disparity=cvCreateImage(cvSize(width,height),8,1);//or IPL_DEPTH_8U
    BwImage imgA(disparity);
    
    for (y=0;y<height;y++){ 
       for (x=0;x<width;x++){
          imgA[y][x]=0;
       }
    }
    
      CvScalar sum;
    //CvScalar s2;
    for (y=0;y<height-N;y++){ //height-N
       for (x=0;x<width-N;x++){//width-N 
          cvSetImageROI(greyLeftImg32, cvRect(x,y,N,N));
          d=0;         
           //initialise localSAD
           for (m=0;m<DSR;m++){localSAD[m]=0;} 
           
           //start matching
           do{ 
             if (x-d>=0){ 
                cvSetImageROI(greyRightImg32, cvRect(x-d,y,N,N)); 
             }else{
                break;
             } 
 
             cvAbsDiff(greyLeftImg32,greyRightImg32,winImg);//absolute difference
             sum=cvSum(winImg);//sum
             localSAD[d]=sum.val[0];//0 means single channel
 
             cvResetImageROI(greyRightImg32); 
             d++;
          }while(d<=DSR);
          
          //to find the best d and store
           imgA[y+offset][x+offset]=getMaxMin(localSAD,DSR,0)*16; //0 means return minimum index
          cvResetImageROI(greyLeftImg32);
       }//x 
       if (y%10==0)cout<<"row="<<y<<" of "<<height<<endl;  
    }//y
    
    cvReleaseImage(&winImg); 
    //cvReleaseImage(&rightWinImg); 
         
    return disparity;
 
 }