Logo Questions Linux Laravel Mysql Ubuntu Git Menu
 

(OpenCV) Fast adjacency matrix computation from watershed

I would like to know if there is a faster way, than what I have done below, to compute the region adjacency matrix from a watershed image.

Input: watershed image with N regions labeled from 1 to N.

Output: adjacency matrix of these N regions.

1. For each region, compute the corresponding mask and put all masks into a vector:

vector<Mat> masks;    
for(int i = 0; i < N; i++ )
{    
// Create the corresponding mask
Mat mask;    
compare(wshed, i+1, mask, CMP_EQ);

// Dilate to overlap the watershed line (border)
dilate(mask, mask, Mat());

// Add to the list of masks
masks.push_back(mask);    
}

2. Define a function to check if two regions are adjacent:

bool areAdjacent(const Mat& mask1, const Mat& mask2)
{
    // Get the overlapping area of the two masks
    Mat m;
    bitwise_and(mask1, mask2, m);

    // Compute the size of the overlapping area
    int size = countNonZero(m);

    // If there are more than 10 (for example) overlapping pixels, then the two regions are adjacent
    return (size > 10);
}

3. Compute the adjacency matrix M: if the i-th region and the j-th region are adjacent, then M[i][j] = M[j][i] =1, otherwise they are equal to 0.

Mat M = Mat::zeros(N, N, CV_8U);
for(int i = 0; i < N-1; i++)
    {
        for(int j = i+1; j < N; j++)
        {
            if(areAdjacent(masks[i], masks[j]))
            {
                M.at<uchar>(i,j) = 1;
                M.at<uchar>(j,i) = 1;
            }
        }
    }
    return M;    
like image 397
f10w Avatar asked Jun 19 '13 08:06

f10w


1 Answers

The following is simple but very fast:

Mat getAdjacencyMatrix(const int* klabels, int width, int height, int K)
/////* Input:
////        - int* klabels: the labeled watershed image (the intensity of the watershed lines is -1)
////        - int K: the number of superpixels (= kmax + 1)
//// * Output:
////        - Mat M: the adjacency matrix (M[i][j] = M[i][j] = 1 if the superpixels i and j are adjacent, and = 0 otherwise)
////*/

{
    /// Create a KxK matrix and initialize to 0
    Mat M(K, K, CV_32S, Scalar(0));

    /// Scan the labeled image
    for(int y=1; y < height-1; y++)
    {
        for(int x=1; x < width-1; x++)
        {
            // Get the label of the current pixel and the ones of its neighbors
            int k = klabels[y*width + x];
            int kleft = klabels[y*width + x - 1];
            int kright = klabels[y*width + x + 1];
            int kup = klabels[(y-1)*width + x];
            int kdown = klabels[(y+1)*width + x];
            if(k != kleft)
            {
                M.at<int>(k,kleft) =1;
                M.at<int>(kleft,k) =1;
            }
            if(k != kright)
            {
                M.at<int>(k,kright) =1;
                M.at<int>(kright,k) =1;
            }
            if(k != kup)
            {
                M.at<int>(k,kup) =1;
                M.at<int>(kup,k) =1;
            }
            if(k != kdown)
            {
                M.at<int>(k,kdown) =1;
                M.at<int>(kdown,k) =1;
            }
        }
    }

    return M;
}
like image 121
f10w Avatar answered Sep 20 '22 15:09

f10w