Architecture: CNN frontend + Bidirectional LSTM + CTC loss
Input: 64×256 grayscale image, normalized to [-1, 1]
Output: Chess move string (e.g., "e4", "Nf3", "O-O", "Qxd7+")
Character set: 28 classes (blank + K, Q, R, B, N, a-h, 1-8, x, +, #, =, O, -)

Model Variants

File	Size	Use Case
`chess_ocr.onnx`	41 MB	Browser inference (ONNX Runtime Web)
`chess_ocr_v47_complete.pth`	~47 MB	Python/PyTorch inference
`best_chess_ocr_v47.pth`	~127 MB	Resume training (includes optimizer state)

Usage (Browser - ONNX Runtime Web)

javascript

import * as ort from 'onnxruntime-web';

// Load model from HuggingFace
const modelUrl = 'https://huggingface.co/GerhardTrippen/chess-ocr-bilstm/resolve/main/chess_ocr.onnx';
const session = await ort.InferenceSession.create(modelUrl);

// Preprocess image to 64x256 grayscale, normalize to [-1, 1]
const inputTensor = new ort.Tensor('float32', imageData, [1, 1, 64, 256]);

// Run inference
const results = await session.run({ input: inputTensor });
const output = results.output.data;

// Decode with CTC (greedy or beam search)
const move = ctcDecode(output, charset);

Usage (Python/PyTorch)

python

import torch
from PIL import Image
import numpy as np

# Load model
checkpoint = torch.load('chess_ocr_v47_complete.pth', map_location='cpu')
model = ChessOCRModel()  # See training notebook for architecture
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

idx_to_char = checkpoint['idx_to_char']

# Preprocess: resize to 256x64, grayscale, normalize
image = Image.open('move_cell.png').convert('L')
image = image.resize((256, 64))
img_array = np.array(image, dtype=np.float32)
img_array = (img_array / 255.0 - 0.5) / 0.5  # Normalize to [-1, 1]
tensor = torch.FloatTensor(img_array).unsqueeze(0).unsqueeze(0)

# Inference
with torch.no_grad():
    output = model(tensor)
    
# CTC greedy decode
predictions = output.argmax(dim=2).squeeze()
move = ''.join([idx_to_char[idx.item()] for idx in predictions if idx.item() != 0])

Training

Trained on the HCS (Handwritten Chess Scoresheet) dataset.

See BiLSTM_v7.ipynb for the complete training code.

Training configuration:

Epochs: 50
Batch size: 32
Learning rate: 0.0005
Optimizer: Adam
Augmentation: 10:1 ratio (rotation ±5°, horizontal scale, shear, vertical shift, brightness/contrast)

Key differences from original Eicher et al. approach:

Trained from scratch (no IAM pretraining)
Modified preprocessing: no horizontal line removal, center-padded alignment
Achieves comparable results: 82.9% MRA vs 82.5% MRA (Eicher 2021)

Acknowledgments & Citations

This model implements the architecture described in:

Eicher, O., Farmer, D., Li, Y., & Majid, N. (2021). "Handwritten Chess Scoresheet Recognition Using a Convolutional BiLSTM Network." ICDAR 2021 Workshops, pp. 245-259.
DOI: 10.1007/978-3-030-86198-8_18

Majid, N., Eicher, O. (2022). "Digitization of Handwritten Chess Scoresheets with a BiLSTM Network." J. Imaging 2022, 8, 31.
DOI: 10.3390/jimaging8020031

Dataset:

Original HCS dataset by Eicher, Farmer, Li, Majid
HuggingFace hosting by Benjamin Kost

License

MIT License - Free for personal and commercial use.

Author

Trained and uploaded by Gerhard Trippen (2025).

Citation

bibtex

@misc{trippen2025chessocr,
  author = {Trippen, Gerhard},
  title = {Chess OCR BiLSTM: Handwritten Chess Scoresheet Recognition},
  year = {2025},
  publisher = {HuggingFace},
  url = {https://huggingface.co/GerhardTrippen/chess-ocr-bilstm}
}