How to get segmented image using k-means clustering ?

Greetings to the members of the community!

I am trying to segment a grayscale image of size m-by-n using k-means clustering ( k = 3, in this case). In this attempt, I am using the normalized values of the pixel coordinates and intensity levels as the coordinates in the feature space to perform k-means clustering.

For e.g. for a pixel located at (x,y) having an intensity of I(x,y) the coordinate of that data point in the feature space is (x/m, y/n, I(x,y)/max(image))

Once the clustering is completed that is the centroid values of the clusters no longer change their coordinates drastically (checked by setting a threshold value) from their positon in the previous iteration, the pixels corresponding to the data points are re-allocated with a new intensity value.

However, when I run the program, the overlayed image doesn't match that obtained from using the in-built function imsegkmeans.The attached code is specific to k = 3. I am not able to figure out what piece I am missing or doing incorrectly.

Please share your valuable suggestions on how do I perform the clustering operation correctly.

Thanks!

Answer 1

[Image Analyst]

I can't follow your code without putting a lot more time into it. There are not enough comments. It looks like you're trying to write your own algorithm for kmeans instead of using the built-in kmeans functions.

In general kmeans is not usually a good algorithm for grayscale segmentation, and certainly not for the cameraman image which has a broad continuum of gray levels.

If you want my demo of using the function kmeans, see attached demo programs.

In addition there is a special kmeans for images: imsegkmeans

https://www.mathworks.com/help/images/ref/imsegkmeans.html?searchHighlight=kmeans%20segmentation&s_tid=srchtitle_support_results_5_kmeans%20segmentation