SEM Image Segmentation/Clustering

Question

I would like to identify clusters of a nano structure. Here is the raw image: http://i.imgur.com/PDEcW4G.png The clusters should look like this: http://i.imgur.com/ZVPaf7p.gif

Prefered tool would be Matlab.

Background information

Overall goal is to distinguish between foreground and background structures. To reconstruct a 3D model of porous media, SEM images are taken from slices, then binarized into solid and pore and finally placed in a row in z-direction. While identifying solid is easy, the pores unfortunately show solid of the subsequent slice. Hence, the idea is to verify which structures change, ie solid of the particular slice, and which do not, being solid of the subsequent slices. As pixel-wise comparison is inaccurate due to (nano) drift of the structure comparing whole clusters promises better recognition.

Suggestions and criticism regarding the overall approach are very welcome!

Answer 1

Image segmentation is a hard problem. Different approaches are suited well to different conditions, and I'm not entirely sure what the "optimum" segmentation you're actually after. If you want to separate "reflecting" from "not reflecting" in the SEM, then you're right -- you're probably better off using the morphological threshold-based operations like you've said in the comment above. You can always use imopen and imclose (in matlab) to morphologically open and close the image (ie connect or shrink structures).

However, in general, no algorithm will work perfectly across a heterogeneous set of images. Your best bet -- ultimately -- is to use the nuclear option and take a machine learning approach with a large amount of training data. Looking at those images, it isn't immediately clear to me what the ideal solution should be -- and that's probably a bad sign.

Good luck!