I am assigned a machine vision task where I have microscopic images of material surfaces. Here is one sample

I am looking to extract regions of interest in such images. To the eye, it’s clear that there is an edge in the image, but edge detection algorithms fail to detect it(fail here means it detects much more than what I need), as it would courtesy of the many varying intensities along the pixels. I have tried various traditional methods to improve the image for edge detection(smoothing, unsharp masking, morphological operations, DCT filtering) but none of them provide satisfactory results. Here is what I believe, the best result from my approach(histogram equalisation, gaussian blur, otsu thresholding, unsharp masking, in the same order)The problem still remains on applying edge detection.

Are there any other approaches to this problem? I mentioned edge as an example, but in general I need any such region of interest. This is a part of my task which is image mosaicing, so I will need to find such features which are the same in 2 images taken by shifting the camera slightly, which I will use to register those images. I am considering CNNs for feature extraction but I doubt any of the available models are suitable for this purpose, so building and training one is something to investigate.