Is this also how some edge detection algorithms work?

@chii I think people usually do edge detection with neural nets nowadays, but not usually with actually doing Fourier transforms; instead, you can just apply a gradient filter to the image and look for local maxima

@liangcyn could you share some pointers of how to use neural nets to for edge detection? I know there are efforts to use NN to extract local descriptors such as KNIFT but that's more like a feature matching problem instead of generating real edges. I guess one could also use a segmentation approach to find edges but that might be too coarse compared with traditional edge detection algorithms. So quite curious about the right way to do it.

@ethanyanjiali, try this. Hopefully it can give the general idea of how people train DL models to detect various features. On slide 41, you will see some extracted line images by deep CNN. Starting with the input image, each layer in the deep learning layers extracts repeated patterns(features) from low dimensions to high dimensions procedurally.

Thanks @movissup, I think this paper did a good job visualizing these low-level features too. Since I was looking for the edge detection task specifically, I went on for some search and these papers are pretty good resources explaining using CNN for edge detection: DeepEdge, HED, and a more recent one here.