Asahi Sato (Master's Program), Graduate School of Engineering, Tokyo University of Agriculture and Technology, is a web application that AI (artificial intelligence) immediately recognizes and converts as phase data when you take a picture of the shogi board and the piece you have with a smartphone and send it to a web server. Developed by Kei Morizumi (2st year master's program).
Image recognition of a shogi board consists of three phases: "recognition of the frame of the shogi board", "recognition of the pieces on the shogi board", and "recognition of the pieces held on the shogi board".Line Segment Detector (LSD) is used for frame recognition of shogi board, deep neural network called Convolutional Neural Network (CNN) is used for piece recognition on the board, and object recognition is used for holding piece recognition. The method YOLO (You Only Look Once) is used.
For learning the recognition model, we used pieces and boards owned by the shogi club and researchers, and images published on the Internet.Since the shooting environment of the user of the application is not constant, we also prepared images with different brightness and size.It is said that various types of pieces such as one-character pieces and two-character pieces, red and black pieces on the back, and special fonts are included.
When this application was tested at a shogi cafe, etc., it achieved a recognition rate of 91% for frame recognition, 99.7% for board recognition, 73.4% for recognition of possessed pieces, and 97.9% for recall.If it cannot be recognized well, a display is displayed so that the photographer can immediately notice it, and the system requests re-shooting.
This application has already been released.The tool that digitizes the situation by image recognition of the shogi board is the first service in Japan and the world, and it is said that it will be a powerful tool to look back on one's game and improve the playing ability.It is also useful for further analysis of shogi, which is said to be unable to search all aspects even with a supercomputer by accumulating a number of aspects of shogi.
