Vision-Based AI Model Solves Sudoku at a Glance

Although “Sudoku“ grid-based number puzzles are no match for today’s artificial intelligence systems, a novel approach to the challenge is trending on GitHub due to its practical integration of computer vision technologies. Users can simply snap a photo of a Sudoku puzzle in a newspaper or playbook and the GUI Smart Sudoku Solver will automatically transfer the image to a computer friendly language, then find and output the answer.

The GUI Sudoku Solver is the brainchild of Neeramitra Reddy, an undergraduate in the department of Computer Science and Engineering at the National Institute of Technology Karnataka in India.

Installation is fairly simple: download Python, create a virtual environment and clone the repository, then connect to the Internet to create the knn.sav file in about five to ten minutes. The Smart Sudoku Solver is then able to run offline with only local files. The modeltype variable is set to the K Nearest Neighbours Algorithm for Recognition by default since this has produced the highest accuracy in experiments. Users however can also choose to set it as a Convolutional Neural Network.

The estimated three percent misrecognition rate of the KNN model can be eliminated via a user interface that enables manual review and editing of any wrong entries before the engine solves the puzzle.

Some 10 years ago, Swedish developer Hans Andersson built an amusing Sudoku solver based on a LEGO mobile robot, which navigated over and detected numbers on Sudoku puzzle printouts using a light sensor, then solved the puzzles using a recursive backtracking algorithm. This new AI Sudoku Solver looks like an upgraded version of the LEGO robot, with faster processing time, improved portability and fewer restrictions.

More information on the model’s Gaussian Blurring stage, noise removal, and other image processing stages are available on the project’s GitHub page.

---

Author: Reina Qi Wan | Editor: Micahel Sarazen