receipt-text-detection_kr-pan_resnet18

receipt-text-detection_kr-pan_resnet18

A lightweight OCR text detection model optimized for Korean receipts and invoices.

🎯 Quick Start

# Load the model (requires pytorch)
import torch

checkpoint = torch.load("epoch-18_step-001957.ckpt", map_location='cpu')

📊 Performance

• Precision: 95.35%
• Recall: 95.72%
• H-Mean: 95.37%

🏗️ Architecture

• Base: ResNet18 encoder
• Decoder: PAN decoder
• Head: DBNet with polygon output
• Framework: PyTorch Lightning

📝 Training Details

• Dataset: Private Korean receipt/invoice corpus
• Epochs: 18/20
• Steps: 1,957
• Optimizer: AdamW (lr=0.001)

⚠️ Important Notes

This model was trained on proprietary, copyrighted receipt data. Publishing is permitted because:

1. The model contains learned representations, not original data
2. No training data is included in this repository
3. The model should be used responsibly in compliance with applicable laws

Users should:

• Not attempt to extract or reverse-engineer training data
• Respect intellectual property rights
• Consider privacy implications when applying to real receipts
• Implement appropriate data protection measures

🔧 Model Specs

• Input: Variable resolution images (recommended 640×480)
• Output: Text region masks and polygon coordinates
• Model Size: ~11.7 MB
• Inference: ~50-100ms (GPU), ~300-500ms (CPU)

📚 Full Documentation

See MODEL_CARD.md for complete model card with:

• Detailed architecture information
• Training configuration
• Usage examples
• Limitations and ethical considerations
• Citation information

📖 Project Reference

For training code and full system context:

• Original: AIBootcamp13/upstageailab-ocr-recsys-competition-ocr-2
• Continuation: Wchoi189/upstageailab-ocr-recsys-competition-ocr-2

📄 License

MIT License