I created a neural network that solves a sudoku. It takes in an image of an incomplete sudoku and gives an image of the completed sudoku as the output.
It uses the Keras MNIST Handwritten Digit Recognition neural network to convert the image into numbers and then uses a backtracking algorithm to solve it.
It's accuracy is the same as that of the Digit Recognition model which is approximately 99%.
Unsolved Sudoku and Solved Sudoku:
I used the following model as a reference: https://levelup.gitconnected.com/solving-a-sudoku-puzzle-using-deep-learning-and-backtracking-algorithm-c6cef475ae3