Dear @Donatella Giuliani.

This is an interesting but complex question.

It all depends on the type of image (medicine, nature ...) and the color system.

Therefore, it is difficult to give an unambiguous answer.

For some images:

I made a translation from RGB to HSV and then worked with channel V.

For other images, I worked on channel R, G, B separately, and then merged the result by mask.

With the best wishes,

See also profile - https://www.researchgate.net/profile/Cecilia_Di_Ruberto/contributions

Dear @Donatella Giuliani.

This is an interesting but complex question.

It all depends on the type of image (medicine, nature ...) and the color system.

Therefore, it is difficult to give an unambiguous answer.

For some images:

I made a translation from RGB to HSV and then worked with channel V.

For other images, I worked on channel R, G, B separately, and then merged the result by mask.

With the best wishes,

See also profile - https://www.researchgate.net/profile/Cecilia_Di_Ruberto/contributions

Hi,

As stated above, I would also recommend the usage of HSV color space. Try to experiment on the generated channels and choose the one which works best for your images; (H,S, or V).

Thanks a lot for your suggestions, have you used also L*a*b color space? have you quantized the color images in a pre-processing step?

No, why do you need to quantise the color?

Dear Abbas,

I need to have essential color information for edge extraction and skeletoniztion, do you agree?

Hi Donatella,

Contrast enhancement is proven to be useful in the context of edge detection. Quantisation, humm, I am not sure to be honest.

p.s. In your original post, you talked about color image segmentation, but now you are after skeletnisation, which are two distinct topics in my eyes :)

K means clustering on HSV color space probably give good result ....

Prof Giuliani  , I have not worked in color space , as far as I understand it is better you look @ e.g MacAdam ellipse  , the coordinates are fractional for you to get exact wavelength range /color (in RGB)

Dear @Donatella Giuliani

In my opinion, the first one you must identify the type of your image.

In further, maybe you can try to mix-and-match RGB and HSV color space. I have combined the Red channel and Saturation channel to segment the blood microscopic image. I also have tried the LAB color space, but it is not suitable for my dataset

Thanks to all of you. With a test image from Berkeley data-set I have used HSV color space with de-correlated images The segmetation results are quite good. Generally which indices do you use for the evaluation of segmentation quality?