Chess Vision

Convert an image of a chessboard into a digital representation

Battle Plan

• LiveView to capture image and store it. File or S3?
• Parse image into array of images of 64 squares
  • Canny edge detection - Rust nif
  • Hough line detection - Rust nif
  • order is important: top left of board will be element 0
  • orientation is not guaranteed; A1 could be any corner of the board
• Convert each of those images into data points by feeding them to an image recognition neural network
  • train NN
  • training/test sets of tactics book images and live board images. Use step 1 to help with this.
  • Use Nx
  • output is array of structs containing { piece_type, piece_color, square_color }
• Convert that data into a FEN string
  • figure out how to determine orientation of the board
  • require picture to be taken at a certain orientation if needed
  • https://en.wikipedia.org/wiki/Forsyth%E2%80%93Edwards_Notation
  • example after 1. e4: "rnbqkbnr/pppppppp/8/8/4P3/8/PPPP1PPP/RNBQKBNR b KQkq e3 0 1"
• Render FEN string as a chess board in the web app UI
  • use chess.js to take a FEN string and display the board accordingly
  • https://github.com/jhlywa/chess.js
  • display FEN string along side original image