k means

I'm trying to perform k means on a 512*512 img of class uint8. Here is my code ab = double(img(:,:)); nrows = 512; ncols = 512; ab = reshape(ab,nrows*ncols,1); [cluster_idx cluster_center] = kmeans(ab,[],'distance','sqEuclidean',... 'emptyaction','Singleton','start',[1,2;2,1]); but this is the error msg I get ??? Subscripted assignment dimension mismatch.

Error in ==> kmeans>batchUpdate at 428 [C(changed,:), m(changed)] = gcentroids(X, idx, changed, distance);

Error in ==> kmeans at 337 converged = batchUpdate(); Please help me figure out where I am going wrong. Thanx in advance.

Answer 1

[Walter Roberson]

Note that your setup code is equivalent to

ab = double(img(:));

You are passing in a single column of data to kmeans(), so kmeans is getting 1D data, but you are trying to initialize the centroids as a pair of two-dimensional points.

[Mahalakshmi]

I get it. So , I tried this based on the demo of k means. In the demo k means is done on an rgb image converted into Lab colorspace. With ab the vector size was 2621441*2. In my code since i'm using an img of class uint8 my vector size was 2621441*1. I'm just wondering if it's possible way to perform kmeans on grayscale images using the inbuilt kmeans function..

[Walter Roberson]

Yes, you can perform kmeans on grayscale images using the inbuilt kmeans function. You must, however, first define what it is you want to cluster on.

Looking through the documentation for kmeans for the version I have and for the latest (2011a) version, I do not see any possibility of using an empty matrix in place of the number of clusters ? Which version are you using?

[Mahalakshmi]

I too use the same version. When I changed the distance measure to cosine or correlation ,this is the error I get..

Error using ==> kmeans at 157

Some points have small relative magnitudes, making them effectively zero.

Either remove those points, or choose a distance other than 'cosine'.