Determining Chess Game State from an Image.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Accurate chess piece identification from images is crucial for automated chess analysis and player improvement.
• Existing computer vision methods struggle with accuracy, large datasets, and adapting to diverse chess sets.

Purpose of the Study:

• To develop a novel, accurate, and adaptable chess recognition system using computer vision and deep learning.
• To create a large-scale, synthesized dataset for training robust chess recognition models.

Main Methods:

• A RANSAC-based algorithm for chessboard localization and transformation to a regular grid.
• Two convolutional neural networks for occupancy mask prediction and piece classification.
• A few-shot transfer learning approach for adapting to unseen chess sets.

Main Results:

• The system achieved a 0.23% per-square error rate, outperforming the state-of-the-art by 28 times.
• The few-shot transfer learning method reached 99.83% per-square accuracy on new chess sets with minimal data.

Conclusions:

• The developed system offers a significant advancement in automated chess recognition.
• The approach demonstrates high accuracy and adaptability, facilitating practical applications for amateur chess players.