At present the heaviest data in the result images are the rotation boundaries. As we know the rotation angle before the images are cut out, one could include a rotation dependent margin and crop the characters square. I expect that's the most significant possible improvement.

It's a mild cop out, but building a custom classifier from the images would definitely give the most reliable match.

This is a fun problem. I had a few nice ideas and these rapidly degraded into a notebook.

The concept is to pick out things looking like digits with morphological tools. One can then prune off items too far from the center, and fit the most extreme digits to a circle, constricting the circle to find the inner rings.

The reality of the data isn't so nice. Not only are the rings at angles to the camera, but this angle changes with the vertical element, so different ellipses are required for each tier. I fear not accounting for this z axis drop within each ring is one of several weakness to the solution. I was quite pleased with it. Notebook attached.

Attachments: