Chess Piece Image Recognition Using Transfer Learning, Simple Neural Network, and Convolutional Neural Network

Online chess has opened up a way for players to sharpen their skills through move analytics. Based on this feature, a support system called chess advisor can be utilized to assist players in a real-time match. However such system doesn't exist within the website itself, rather an involvement of 3rd party software is required. These kinds of software usually need assistance for input. This paper aims to create a system that is able to recognize chessboard state in real-time. We have proposed a chess piece image recognition system using transfer learning using ResNet50, a simple neural network, and a convolutional neural network (CNN). We collected our own 693 2D chess piece images dataset and performed data preprocessing to standardize the image dimensions. The proposed system detects and converts individual positions of each piece into a computer-readable output. This output can be used to facilitate the chess advisor in determining the best move. The transfer learning approach utilizing the ResNet50 model exhibited superior performance with an accuracy of 95.65%. This outperformed alternative methods such as the simple neural network and the CNN. The ResNet50 model's advantage stemmed from its pre-training on the ImageNet dataset, enabling effective pattern generalization. In contrast, the simple neural network and CNN achieved lower accuracies of 73.4% and 85%. These models encountered difficulties due to limited dataset quantity and quality, as well as the absence of regularization techniques. Nevertheless, the CNN displayed better consistency in generalizing patterns compared to the simple neural network.