You can try selecting a few, and then use ‘show features’ to see how they look for a bit of the image as an ImageJ stack.
An overview of the features currently:
- Gaussian: a smoothed version of the original image
- Laplacian of Gaussian: good for ‘blob-like’ things
- Weighted deviation: intended to reflect ‘local variation’… zero in flat parts of the image, higher if a lot if happening
- Gradient magnitude: good if you want to detect edges
- Hessian determinant: good for ‘blob-like’ things; this can be more specific than Laplacian of Gaussian (which can pick up other kinds of edges), but weirdly doesn’t distinguish between bright & dark blobs… so you might need both
- Hessian eigenvalues (middle only exists for z-stacks): mostly for lines/ridges
In general, I stick with Gaussian and select a few scales; this gives the ‘original’ image smoothed to varying degrees. I find it I add too many features, the classification often falls apart… but with Gaussian it is more predictable and easier to interpret when it makes mistakes. The caveat is that it is really just using local color information and not much else.
I don’t feel the features have been tamed yet, and I’m still on the lookout for better ones to include in the pixel classifier.