Takes in sudoku images as input with handwritten digits and classifies the digits and identifies box edges.
It then solves these sudokus and reprojects the solutions on the original image.
I have trained a CNN from the MNIST database to help recognize the digits with an accuracy of 99.06%
This program succeeds in identifying the rows and columns from the sudoku puzzle and recognizing the digits.
Due to the use of a custom CNN, the accuracy is not high enough to predict every digit of the puzzle correctly.
Due to incorrect predictions of even 1 digit, the sudoku will then not have a solution.
Thus, the sudoku solver does not work in practical use, but if used with an already trained model of much higher accuracy, one can get the solution to the sudoku.
In future updates, I will improve the accuracy of the CNN and thus solve the lack of practical application of this sudoku solver.