Chess Vision: Real-Time Board Recognition & Analysis

A computer vision application that recognizes chess positions from screenshots and provides engine analysis. Combines deep learning, computer vision, and chess engine integration into a cross-platform pipeline.

Architecture

Screen Capture → Chessboard Detection → Grid Segmentation → CNN Inference → FEN Generation → Stockfish Analysis

Input: Cross-platform screen capture (mss on Linux/Mac, win32gui on Windows) with automatic browser window detection.

Vision: Chessboard detection via OpenCV (Canny edge detection, contour analysis), 8×8 grid segmentation with coordinate mapping (A1-H8), image preprocessing (100×100 normalization).

Inference: Custom CNN model (TensorFlow/Keras) for 13-class piece classification (6 black pieces, 6 white pieces, empty square).

Engine: FEN generation from predictions using python-chess, Stockfish integration for evaluation and best move calculation.

Setup

Prerequisites

• Python 3.10+
• Stockfish binary (download)

Installation

1. Clone and setup environment:

   git clone <repository-url>
   cd chess-vision-tf
   python -m venv venv
   source venv/bin/activate  # Windows: venv\Scripts\activate
   pip install -r requirements.txt

2. Install Stockfish binary:

   • Place executable in engines/ directory:
     • Windows: engines/stockfish.exe
     • Linux/Mac: engines/stockfish
   • Or ensure Stockfish is in system PATH

3. Place trained model at models/model.h5

Platform Notes

Windows: Automatically detects and captures browser windows.

Linux/Mac: Captures primary monitor by default. For window-specific capture, install xdotool or wmctrl.

Usage

python main.py

Two input methods:

• Screenshot Analysis: Captures screen, detects chessboard, classifies pieces, opens analysis window
• FEN Input: Directly load a position using Forsyth-Edwards Notation

License

MIT License - see LICENSE file for details.