When playing Go on a physical board, it can be a pain to record the game to learn from later. Particularly, the .sgf tree format provides an easy way to view, send and make hypothetical changes to a completed game. This project converts webcam pictures of a game into a digital reprentation of each turn, then convertes those states into a .sgf file.

This project takes an image of a Go board as input and outputs a matrix representation of the game's state.

The edges of the Go board are detected in the inputted image, which is then transformed to a square, top-down perspective. The image is then segmented by each square, then fed into a Convolutional Neural Network (CNN) trained to classify each sub-image as containing a white piece, black piece or no piece. The classification for each position is stored in an array, and the array representation is converted to a .sgf representation. This is then written and can be viewed by most types of Go software.

Taking multiple game states and ordering them into a tree structure.

Cleaning up the pipeline into a user-friendly api that can be accessed by a frontend