[SOLVED] Generic detection of subimages in images with openCV

Generic detection of subimages in images with openCV

Disclaimer: I'm a computer vision rookie.

I have seen a lot of stack overflow posts of how to find a specific sub-image in a larger image.

My usecase is a bit different since I don't want it to be specific and I'm not sure how I can do this (if it's even possible, but I have a feeling it should).

I have a large datasets of images, of sometimes, some of this images are a combination of two or more other images of the dataset. I'd like to automatically crop theses "combinations" to isolate the sub-images.

So the tasks would be to process each image of the dataset, and check if there are abnormal boundaries that could mean the image is a combination.

Example using great stock images:

What I've tried:

I've seen that houghs transform could be used for line detection in images but I'm couldn't achieve anything using this.

Solution

Use cv.Sobel or other derivative kernel (with cv.filter2D) to find edges
Sum along pixel columns to score each for "edginess". np.sum or np.mean do that.
Some thresholding and np.argsort and indexing to find best candidates.
cut/split (take slices out of array)

edge map:

edgy plot:

slices to take:

array([[   0,  578],
       [ 578, 1135],
       [1135, 1136]], dtype=int64)

Pictures:

complete notebook: https://gist.github.com/crackwitz/e1ba1ce7a6fba446288275d91f66261c