A Python program that uses computer vision and machine learning to solve any Sudoku puzzle from an image. Digit detection is done with a CNN trained on the MNIST dataset - the saved model and the script to train it is located in this repository.
This demo showcases an solution of the program on this initial image:
The solution:
- After turning the image grayscale, we blur the it to smooth out noise. Makes extracting of lines easier.
- Threshold the image leaving only blacks and whites.
- The image is inverted to help identify black borders.
- Dilation is applied to fill in gaps in lines and digits
- The largest contour (border of puzzle) is found in order to find the Sudoku in the image
- The image is warped and a grid is inferred
- Each digit is extraced from the grid and identified using a CNN trained on the MNIST dataset
- The digits get passed so a backtracking algorithm solving the puzzle which is the presented on the warped version of the original image
