Color segmentation problem using matlab

Good evening All, My problem is a regular problem: extracting features of similar color (segmentation) from a scanned image. >> I have tried Kmeans segmentation I require something more precise close to exact to extract the colors from the scanned images. Are there any other means but Kmeans segmentation that may solve my problem?

>> Is/are there any ways to use some reference areas and segment the image in frequency domain?

Thank you

Answer 1

[Image Analyst]

You will find a treasure trove of color segmentation demos in my File Exchange: http://www.mathworks.com/matlabcentral/fileexchange/?term=authorid%3A31862.

I have no idea what you're thinking of when you talk about segmenting the colors in the frequency domain. That does not make any sense at all to me. Care to explain how that could happen?

[Image Analyst]

Siamack's "Answer" moved here because it's not an answer, it's a reply to my answer:

"Yes absolutely, what I meant was if there were any chances one could take a sample color from a designated group of pixels and use their fourier transform to find a significant measure of reference by means of which could extract every pixel (or region) similar (in frequency spectrum) to the reference area from the whole image. I have scanned through the page you have introduced and have picked up some helpful codes but unfortunately non of them enables me to reach the precision I am looking for. There seems to be an awful amount of cross-talk between the regions (I am scanning actual emissive regions). Thanks btw for taking your time to help me out. Hope I made some sense with what I wrote in response."

[Image Analyst]

I still have no idea what you mean. When you say frequency, you mean spatial frequency, right? Like how smooth or detailed the image is? You don't mean frequency like frequency of light, which is color, like the wavelength of light, do you?

I don't see how you can take "a sample color from a designated group of pixels and use their fourier transform to find a significant measure of reference". I just don't know what that means. If you have, say an aerial photo of blue lakes, and then you take a sample of one lake, then that one lake will have some spatial frequencies in it. This means it has a texture, like the water is smooth, or rippled. Now I don't know how you are going to take the fft of some irregular part of a lake and find other blue things in the image, like sky, cars, etc. If you did a texture segmentation you could find everything that was a similar texture to the lake: smooth or rippled, but that would get all colors. You'd still have to do a color segmentation after you did the texture segmentation. The spatial frequency spectrum just doesn't have any color information in it. All it has is info on the detail in the image - how fast the image varies in intensity, not what color it is.

[Siamack]

Hi there, Sorry for the delay, No and Yes re: if im looking at the frequency of the incoming light No altho I find it quite interesting except for my wits do not allow me to further understand if there could be a way to use that as means to analyze the sample in frequency domain. My assumption is based on 1D frequency domain analysis which simply finds (reconstructs or preserves) certain frequency of change in the input signal through suppression or amplification. That translates into the spatial frequency of course and I know that it may not contribute to color due to its correlative change with respect to texture but looking at an image as a matrix with a surface represented by each numerical value does give an idea of a topological consistency throughout the image (in my experiment's case especially) over the areas which are emitting the same color. Hence suppressing any other texture (including the background) does my required job: segments the image with the precision I am looking for....

P.S How does the frequency of the incident light might play a role in frequency domain analysis? do you have an idea where to start diggin around it?

Regards, Siamack

[Image Analyst]

Electromagnetic frequency of the light, and spatial frequency of the image are pretty much unrelated (ignoring any diffraction effects that you may get if you're looking at a field of view or things in the optical path that are small enough to cause diffraction).

Are you saying that not one of the 3 different color segmentation methods I showed you in my File Exchange can be properly adapted to segment your image? If that's the case you'd have to post/attach your image to get further advice.