How to merge super pixels to get bigger region segmentation

I have a input color image. I have already segment the image using super pixel function and divide the region into 500 superpixels. Now i can see that boundary of each objects are properly segmented. So i want to merge adjacent superpixels using some homogenity. How to merge superpixels to get large region based on some homogenous chracateristics. please help me out.

I have attached a image with superpixels based segmentation. Now how to merge adjacent superpixels based on some homogenity.

Answer 1

[Asma Semche]

I am, kind of, doing the same thing, I want to use the Bhattacharyya distance in order to merge segments.

The thing is, I can't find a way, yet, to regroup pixel intensity of the borders in a commun matrix.

I'd be happy if we can solve your problem and mine, together.

[MB11]

Even I need to do the same with Bhattarcharya coefficient. Have either of you found any solution?

Answer 2

[Image Analyst]

Not so easy, and not unique. What if region A is close enough to region B to merge, and then what if region B is close enough to another adjacent region C to merge. But what if region A is not similar enough to region C to be considered the same. If you merged B with both A and C you're going to have dissimilar regions merged into the same region. Do you just merge ONE of the regions into B -- the closest one? But then I can see that the output you'd get would depend on the ORDER that you merge the regions. If you started merging with one region, you may end up with a completely different output than if you started with some other region. What do you plan to do in that case? Anyway, why do you want to merge them anyway? Maybe you can just use multithresh() to control what similar regions get grouped together. What is your use case?

[TAUSEEF KHAN]

i want to divide the whole image based on iso-intensity regions based on some tolerance factor or called threshold. if i choose small tolerance factor then, more regions will be generated and vice versa. Final output may seems like, foreground objects will be segmented in a better way

[MatlabUser]

do you have an idea how to merge the superpixels if the threshold is fixed ?

[Image Analyst]

I just don't see a way in superpixels() to specify any parameters except for the NUMBER of superpixels. Why do you want superpixels anyway? If you just want to label similar areas you could use linear discriminant analysis, like with the attached demo. This lets you trace out training areas that you think are roughly the same color, and then it classifies every pixel in the image as one of your training colors.

See attached demo.

[MatlabUser]

How can I use linear discriminant analysis if the segmentation is unsupervised, I wanted to be automatic

[Image Analyst]

Try imsegkmeans() or rgb2ind(), but you have to know how many different color classes you want.