I am learning image processing using Matlab currently. I have RGB image from a DSLR camera. I would like to segment image area which has slightly different color and intensity. Which color space is the best for this task? What is the reason? Please also suggest me which segmentation techniques better for this?
Otsu's method can give you automatic thresholds for segmentation.
As you might imagine, there are several techniques for segmentation (such as graph cuts, normalized cuts, belief propagation, watershed etc.),
you can also use machine learning based techniques (such as semantic segmentation using FCNs) to learn the best segmentation, if you feed the algorithm some already segmented images for your specific problem
what's best? it depends on the type of image you have, try the above mentioned techniques and pick one.
do you want semi-automatic (manual intervention) or full automatic techniques?
Thank you for your answer Vikas,
I still learning and considering whether I will use semi-automatic (supervised?) or full automatic techniques. What do I need to consider in selecting the appropriate technique? Should I try all of them?
Hi Dimas!
As rightly stated by Vikas, it very much depends on the type of images and the nature of the ROI you are after.
Sticking to your question about the color spaces, if your ROI has a clear contrast over the background in terms of color then I would recommend doing color transformation (e.g., HSV, YCbCr, L*a*b* ). Note that such transformation will allow you to take advantage of color separation in the chrominance space without interference of the luminance (a.k.a. brightness). Also note that color-based segmentation has normally low time complexity (not considering the non-linear transformation of the HSV).
Face segmentation is an example where human skin tone tends to cluster around certain region in the red chrominance. As for the potato defect detection, you might wish to experiment on the above color spaces to see which one helps addressing your problem better.
Hope that helps in some sense.
Thank you Abbas,
There is some area that could be segmented by color, however, there is also some area that has a similar color but different luminosity. I want to segment all of that area. That is why I confuse where to start, maybe I will try to segment it first for all area and then feature extraction or combine different segmentation method.
If you want a segmentation by hue (seems to be your intention) HSV should work best. Anyway, in very dark or very bright regions, HSV might give results you'd consider 'false' at the first glance. But mathematically, these will be about the best results you can 'wring' from the individual image. . . . When it comes to segmentation, the human eye-brain combination is to some extend still superior.
Regards
As everyone said it depends on your images.
My suggestion is to try several colorspaces, inspect their channels individually with a few sample images to select the channels with higher probability of success. Then try some methods over that channels and you may get the best result from a combination of channels from different colorspaces.
There is no magic récipe, you have to try several hipothesys
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing