Takes an uncontrolled image of a sudoku puzzle, identifies the location, reads the puzzle, and solves it.
This example was originally put together for the UK MATLAB Expo 2018, for a talk entitled Computer Vision and Image processing with MATLAB (video, blog post). It is intended to demonstrate the use of a combination of deep learning and image procesing to solve a computer vision problem.
- Get a copy of the code either by cloning the repository or downloading a .zip
- Run the example live script getting_started.mlx, or see the usage section below
Broadly, the algorithm is divided into four distinct steps:
- Find the sudoku puzzle in an image using deep learning (sematic segmentation)
- Extracts each of the 81 number boxes in the puzzle using image processing.
- Read the number contained in each box using deep learning.
- Solve the puzzle using opimisation.
For more details see the original Expo talk.
Input:
Result:
- In MATLAB set the top level directory as the working directory then runn
install()to add the required folders to the MATLAB path. - Run
setupDataGitHub()to fetch the required training data from GitHub. The data is ~70 MB, downloading and extracting this can take a few minutes. - Run
sudoku.trainSemanticSegmentation(). This will train the semantic segmentation network and save the trained network in themodels/folder. - Run
sudoku.trainNumberNetwork(). This will train the number classification network and save the trained network in themodels/folder. - Once both networks have been trained you can process an image as follows:
im = imread("docs/example.jpg");
% reduce the size of the example image
im = imresize(im, 0.5);
solver = sudoku.PuzzleSolver();
solution = solver.process(im)For reference the training curves for the two networks should look as follows:
Classfication training
Segmentation training
By the end of training, the segmentation results should look something like this:
Please file any bug reports or questions as GitHub issues.
Copyright 2018-2019 The MathWorks, Inc.