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.

CodePudding user response：

• 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 (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