MHRN-ST: Go Board Recognition With Multi-stage Highlight Removal Network Based On Swin Transformer

Artificial intelligence has generally surpassed humans in the game of Go, and the actual human-machine gaming environment is often more complex. In this environment, issues such as highlights, piece misplacement, and occlusions on the image can severely affect board recognition and reduce the intelligence level of the robot player. This paper proposes a multi-stage highlight removal network based on Swin Transformer, which can remove highlights on the image and recognize Go game information. Firstly, the image is processed by a multi-task detection network based on YOLOv5 to obtain highlight information. Then, the image is enhanced with the highlight information and input into the highlight removal network based on Swin Transformer for highlight removal. The output image is further processed by the detection network to obtain board information. We collected a total of 4242 images from publicly available datasets to train the detection network, which to some extent solves the problem of piece displacement. Additionally, we collected 1096 pairs of images for training the highlight removal network. Finally, the effectiveness of the proposed method was tested on a test set containing 56 images, achieving a recognition accuracy of 95.67%.