OpenCV - 如何在灰度图像上应用 Kmeans？

Question

I am trying to cluster a grayscale image using Kmeans.

First, I have a question:

Is Kmeans the best way to cluster a Mat or are there newer more efficient approaches?

Second, when I try this:

Mat degrees = imread("an image" , IMREAD_GRAYSCALE);
const unsigned int singleLineSize = degrees.rows * degrees.cols;
Mat data = degrees.reshape(1, singleLineSize);
data.convertTo(data, CV_32F);
std::vector<int> labels;
cv::Mat1f colors;
cv::kmeans(data, 3, labels, cv::TermCriteria(cv::TermCriteria::EPS + cv::TermCriteria::COUNT, 10, 1.), 2, cv::KMEANS_PP_CENTERS, colors);
for (unsigned int i = 0; i < singleLineSize; i++) {
    data.at<float>(i) = colors(labels[i]);
}

Mat outputImage = data.reshape(1, degrees.rows);
outputImage.convertTo(outputImage, CV_8U);
imshow("outputImage", outputImage);

The result ( outputImage ) is empty.

When I try to multiply colors in the for loop like data.at<float>(i) = 255 * colors(labels[i]); I get this error:

Unhandled exception : Integer division by zero.

How can I cluster a grayscale image properly?

Answer 1

It looks to me that you are wrongly parsing the labels and colors info to your output matrix.

K-means returns this info:

• Labels - This is an int matrix with all the cluster labels. It is a "column" matrix of size TotalImagePixels x 1.

• Centers - This what you refer to as "Colors". This is a float matrix that contains the cluster centers. The matrix is of size NumberOfClusters x featureMean.

In this case, as you are using BGR pixels as "features" consider that Centers has 3 columns: One mean for the B channel, one mean for the G channel and finally, a mean for the R channel.

So, basically you loop through the (plain) label matrix, retrieve the label, use this value as index in the Centers matrix to retrieve the 3 colors.

One way to do this is as follows, using the auto data specifier and looping through the input image instead (that way we can index each input label easier):

    //prepare an empty output matrix
    cv::Mat outputImage( inputImage.size(), inputImage.type() );

    //loop thru the input image rows...
    for( int row = 0; row != inputImage.rows; ++row ){

        //obtain a pointer to the beginning of the row
        //alt: uchar* outputImageBegin = outputImage.ptr<uchar>(row);            
        auto outputImageBegin = outputImage.ptr<uchar>(row);

        //obtain a pointer to the end of the row
        auto outputImageEnd = outputImageBegin + outputImage.cols * 3;

        //obtain a pointer to the label:
        auto labels_ptr = labels.ptr<int>(row * inputImage.cols);

        //while the end of the image hasn't been reached...
        while( outputImageBegin != outputImageEnd ){

            //current label index:
            int const cluster_idx = *labels_ptr;

            //get the center of that index:
            auto centers_ptr = centers.ptr<float>(cluster_idx);

            //we got an implicit VEC3B vector, we must map the BGR items to the
            //output mat:
            clusteredImageBegin[0] = centers_ptr[0];
            clusteredImageBegin[1] = centers_ptr[1];
            clusteredImageBegin[2] = centers_ptr[2];

            //increase the row "iterator" of our matrices:
            clusteredImageBegin += 3; ++labels_ptr;
        }
    }