Image recognition + puzzle solving

(1) There are automated solvers for many number puzzles, say something like Sudoku. These start from a digitised version of the puzzle.

(2) There are image/digit recognition algorithms which can digitise a puzzle from a photo.

(3) There are combined systems which digitise a puzzle from an image (2), then solve it (1).

However image recognition algorithms aren’t perfect. Especially if we start from bad quality images or hand-written digits. Instead of treating the image recognition phase as an opaque process, we can treat it as a transparent one.

(4) A combined system which is able to inform the digit recognition by trying to solve the puzzle!

In this project the goal is to evaluate the difference between doing (3) and (4). We don’t have to (only) do Sudoku. I am happy to discuss the details with the interested student.

Background

CS4402 is not a prerequisite but it would be helpful.

Familiarity with using at least one Machine Learning toolkit is strongly preferred.