Go Game Capture and Reconstruction of Missing Moves Using Deep Learning Techniques

Recording and broadcasting Go games in real time remains a challenging task, as traditional methods rely on electronic boards or manual transcription, both of which have limitations in cost, convenience, and accuracy. This paper presents an approach for automatically detecting board states and reconstructing move sequences using computer vision and machine learning techniques. A Convolutional Neural Network (CNN) and a Random Forest (RF) classifier were trained to recognize individual board intersections, demonstrating robust performance across varying lighting conditions and board configurations. To determine the move order when multiple stones are detected simultaneously due to disrupted images, KataGo was used to evaluate all possible variations of the new moves, based on score lead differences. The results indicate that the actual game sequence tends to have a lower mean score fluctuation compared to randomized move orders, supporting the hypothesis that KataGo can provide reasonable reconstructions of move sequences. Evaluations using recorded games from strong amateur and professional players demonstrated promising accuracy in move detection and sequence inference. The implementation is open source, serving as a foundation for further research and development toward automated Go game transcription and real-time broadcasting.