In this little project I trained a CNN with Tensorflow on the MNIST dataset to be able to recognize digits. The recognized digits will be arranged into a sudoku board and a backtracking algorithm tries to solve it. Once a solution is found a picture of the solved sudoku is displayed.
You can find the CNN in the model.py. Using the train.py with -dl or --download arguments will download the official MNIST dataset and save it in binary form into the data folder. If you already have them downloaded, you can skip the arguments.
The extractor.py has the methods to extract the sudoku board and the digits in the cells.
The solver.py contains the backtracking algorithm.
The main.py is the entrypoint of this tool. The models and the input images are hardcoded.
The code is based on this tutorial