rajesh-1902/greek-htr-sentence-model

Greek Handwritten Text Recognition (HTR) Model

Model Description

This model recognizes Greek handwritten text at the sentence level using a CRNN+CTC architecture.

Architecture:

• 5-layer CNN for feature extraction
• 2-layer Bidirectional LSTM for sequence modeling
• CTC loss with space-awareness improvements

Training Dataset: rithwikn/greek_combined_dataset

Image Size: 1024 × 96 pixels

Special Features:

• Space-weighted CTC loss
• Trainable space logit bias
• Column-wise normalization
• Adaptive thresholding

Usage

# Load the model
from Model import Model, DecoderType

# Load character list
with open('charList.txt', 'r', encoding='utf-8') as f:
    charList = f.read()

# Initialize model
model = Model(charList, DecoderType.BeamSearch, mustRestore=True)

# Run inference
# (See main.py for full inference example)

Training Details

• Batch Size: 8
• Learning Rate: 0.0001 (Adam optimizer)
• Early Stopping: 10 epochs patience
• Space Weight: 3.0

Performance

(Check accuracy.txt for latest validation metrics)

Citation

If you use this model, please cite:

@misc{greek-htr-2024,
  author = {rajesh-1902},
  title = {Greek Handwritten Text Recognition Model},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/rajesh-1902/greek-htr-sentence-model}
}