S1M0N38/chess-cv

Chess CV

Lightweight CNNs (156k parameters each) for chess board analysis from 32×32 pixel square images. The project includes three specialized models trained on synthetic data from chess.com/lichess boards, piece sets, arrow overlays, and centering variations:

• Pieces Model: Classifies 13 classes (6 white pieces, 6 black pieces, empty squares) for board state recognition and FEN generation
• Arrows Model: Classifies 49 classes representing arrow overlay patterns for detecting chess analysis annotations
• Snap Model: Classifies 2 classes (centered vs off-centered pieces) for automated board analysis and piece positioning validation

Quick Start

pip install chess-cv

from chess_cv import load_bundled_model

# Load pre-trained models (weights included in package)
pieces_model = load_bundled_model('pieces')
arrows_model = load_bundled_model('arrows')
snap_model = load_bundled_model('snap')

# Make predictions
piece_predictions = pieces_model(image_tensor)
arrow_predictions = arrows_model(image_tensor)
snap_predictions = snap_model(image_tensor)

Alternative: Load latest version from Hugging Face Hub

from huggingface_hub import hf_hub_download
from chess_cv.model import SimpleCNN
import mlx.core as mx

# Download latest weights from Hugging Face
model_path = hf_hub_download(repo_id="S1M0N38/chess-cv", filename="pieces.safetensors")
model = SimpleCNN(num_classes=13)
weights = mx.load(str(model_path))
model.load_weights(list(weights.items()))
model.eval()

Models

This repository contains three specialized models for chess board analysis:

♟️ Pieces Model (pieces.safetensors)

Overview:

The pieces model classifies chess square images into 13 classes: 6 white pieces (wP, wN, wB, wR, wQ, wK), 6 black pieces (bP, bN, bB, bR, bQ, bK), and empty squares (xx). This model is designed for board state recognition and FEN generation from chess board images.

Training:

• Architecture: SimpleCNN (156k parameters)
• Input: 32×32px RGB square images
• Data: ~93,000 synthetic images from 55 board styles × 64 piece sets
• Augmentation: Aggressive augmentation with arrow overlays (80%), highlight overlays (25%), move overlays (50%), random crops, horizontal flips, color jitter, rotation (±10°), and Gaussian noise
• Optimizer: AdamW (weight_decay=0.001) with LR scheduler (warmup + cosine decay: 0→0.001→1e-5)
• Training: 1000 epochs, batch size 64

Performance:

Dataset	Accuracy	F1-Score (Macro)
Test Data	99.90%	99.90%
S1M0N38/chess-cv-openboard *	-	98.56%
S1M0N38/chess-cv-chessvision *	-	92.28%

* Dataset with unbalanced class distribution (e.g. many more samples for empty square class), so accuracy is not representative.

---

↗ Arrows Model (arrows.safetensors)

Overview:

The arrows model classifies chess square images into 49 classes representing different arrow overlay patterns: 20 arrow heads, 12 arrow tails, 8 middle segments (for straight and diagonal arrows), 4 corner pieces (for knight-move arrows), and empty squares (xx). This model enables detection and reconstruction of arrow annotations commonly used in chess analysis interfaces. The NSEW naming convention (North/South/East/West) indicates arrow orientation and direction.

Training:

• Architecture: SimpleCNN (156k parameters, same as pieces model)
• Input: 32×32px RGB square images
• Data: 4.5M synthetic images from 55 board styles × arrow overlays (3.14M train, ~672K val, ~672K test)
• Augmentation: Conservative augmentation with highlight overlays (25%), move overlays (50%), random crops, and minimal color jitter/noise. No horizontal flips to preserve arrow directionality
• Optimizer: AdamW (lr=0.0005, weight_decay=0.00005)
• Training: 20 epochs, batch size 128

Performance:

Dataset	Accuracy	F1-Score (Macro)
Test Data (synthetic)	99.99%	99.99%

The arrows model is optimized for detecting directional annotations while maintaining spatial consistency across the board.

Limitation: Classification accuracy degrades when multiple arrow components overlap in a single square.

---

📐 Snap Model (snap.safetensors)

Overview:

The snap model classifies chess square images into 2 classes: centered ("ok") and off-centered ("bad") pieces. This model is designed for automated board analysis and piece positioning validation, helping ensure proper piece placement in digital chess interfaces and automated analysis systems.

Training:

• Architecture: SimpleCNN (156k parameters)
• Input: 32×32px RGB square images
• Data: 1.4M synthetic images from centered and off-centered piece positions (985,960 train, ~211,574 validate, ~210,466 test)
• Augmentation: Conservative augmentation with arrow overlays (50%), highlight overlays (20%), move overlays (50%), mouse overlays (80%), horizontal flips (50%), color jitter, and Gaussian noise. No rotation or geometric transformations to preserve centering semantics
• Optimizer: AdamW (weight_decay=0.001) with LR scheduler (warmup + cosine decay: 0→0.001→1e-5)
• Training: 200 epochs, batch size 64

Performance:

Dataset	Accuracy	F1-Score (Macro)
Test Data (synthetic)	99.93%	99.93%

The snap model is optimized for detecting piece centering issues while maintaining robustness to various board styles and visual conditions.

Use Cases:

• Automated board state validation
• Piece positioning quality control
• Chess interface usability testing
• Digital chess board quality assurance

Training Your Own Model

To train or evaluate a model yourself:

git clone https://github.com/S1M0N38/chess-cv.git
cd chess-cv
uv sync --all-extras

# Generate training data for a specific model
chess-cv preprocessing pieces  # or 'arrows' or 'snap'

# Train model
chess-cv train pieces  # or 'arrows' or 'snap'

# Evaluate model
chess-cv test pieces  # or 'arrows' or 'snap'

See the Setup Guide and Train and Evaluate for detailed instructions on data generation, training configuration, and evaluation.

Limitations

• Requires precisely cropped 32×32 pixel square images (no board detection)
• Trained on synthetic data; may not generalize to real-world photos
• Not suitable for non-standard piece designs
• Optimized for Apple Silicon (slower on CPU)

For detailed documentation, architecture details, and advanced usage, see the full documentation.

Citation

@software{bertolotto2025chesscv,
  author = {Bertolotto, Simone},
  title = {{Chess CV}},
  url = {https://github.com/S1M0N38/chess-cv},
  year = {2025}
}

Repo: github.com/S1M0N38/chess-cv • PyPI: pypi.org/project/chess-cv