Struggling with using an OMR to detect checkboxes

Question

Good morning! Right now, I'm working on a project that takes a paper form and converts it into an Excel spreadsheet. The unfortunate problem I'm working to solve is actually detecting the checkboxes and when they're filled in. So far, I've gotten the code to separate the questions into chunks and remove most of the clutter (i.e, the text of the questions, save one or two 'floater' words).

Does anybody know of a good API I could just clunk into this before having to attempt to train an AI myself? Or any good solutions?

So far, I've tried looking for circle-esque contours and counting them. Unfortunately, due to the nature of the form , the contours of the circles usually aren't detected. They're fairly small as well, only ~8px * 8px.

I've also tried looking around where they should be (again, due to it being printed-> scanned, they aren't in the same place every time). This leads me to think my best bet is to use an OMR that's already trained on what an empty circle vs check mark looks like, but I've got no idea where to go from there.

Below are the binarized images of what the form looks like.

im1 im2 im3