Calculate GLCM feature vector for each pixel of an image

Dear all,

I used graycomatrix and graycoprops functions to obtain texture feature which are single values of each feature. Now I want to calculate the feature vector of each pixel in the image based on the feature values obtained. Could anyone provide me any solution for that ?

[Rik]

As I understand it, a GLCM is a larger scale feature, so it is meaningless to have it for each pixel. A bit like asking for a function that returns the mean pixel value for each pixel.

So what is it that you want as a result?

[Ynne]

Thanks Rik for your comment,

As a result I want to use the obtained feature vector as an input to a segmentation step in order to enhance the quality of the segmentation and eliminate if exists irrelevant features. Perhaps I didn't understand how should I deal with this

[Image Analyst]

I agree with Rik. Since you'll have only 8 entries in the GLCM, it's very doubtful it will be useful to you. Perhaps if you give us your image and existing segmentation algorithm we can suggest ways to "enhance the quality of the segmentation and eliminate if exists irrelevant features".

[Ynne]

the segmentation algorithm is k-means, independant from the input image which contains different regions

Answer 1

[Image Analyst]

I doubt glcm stats on a per pixel basis will be a useful feature for you. Attached is an example of GLCM on the entire image. Basically you'd have to do that inside a nested for loop where you scan over every pixel. I really think it's literally a waste of time. Why don't you just post your image and tell us what you want to find, and why you think this algorithm will work for that?

[Ynne]

Thanks for your response.

Attached is an example of an image which my final objective is to segment it into regions using kmeans based on feature extraction

[Image Analyst]

kmeans is lousy for color segmentation. I'm attaching my kmeans demo for segmenting color images so you can try it for yourself and see that it's not great.

One problem with it is that you have to specify how many clusters there are in advance, and for an arbitrary photo, that's not known. So if you wanted to find grass, sky, and horses, and you say there are 3 color classes, that's fine (well actually not) if there are horses there, but what if the scene doesn't have any horses? Don't you want to know that no horses are in the scene? But you won't be able to because you're saying 3 colors and it will have to take the remaining colors and make 3 clusters out of them. So you will have some area for color class #3, but it's NOT a horse that was found!

Better is to use something else, like DEM: https://www.mathworks.com/matlabcentral/fileexchange/37197-dem--diffused-expectation-maximisation-for-image-segmentation

Or an example-based approach like discriminant analysis, done by the classify() function in the other attached demo.

[Ynne]

Thanks a lot. But when using DEM method we are not considering texture features. Is that correct ? If so, how can we integrate feature selection in this segmentation approach ?