Thareah/ocr_khmer_lite_train

Khmer OCR Recognition Model

🇰🇭 High-accuracy OCR model for Khmer text recognition using PaddleOCR framework

Model Overview

This CRNN-based OCR model is specifically trained for Khmer (Cambodian) text recognition, achieving 98.45% accuracy on validation data. The model is optimized for recognizing short text segments (3-5 words) commonly found in documents, signs, and printed materials.

🏗️ Model Architecture

• Framework: PaddleOCR 2.7+
• Algorithm: CRNN (Convolutional Recurrent Neural Network)
• Backbone: ResNet34
• Neck: SequenceEncoder with RNN (hidden_size: 256)
• Head: CTCHead with CTC Loss
• Input Shape: [3, 32, 320] (channels, height, width)
• Max Text Length: 25 characters

📝 Supported Characters

The model recognizes 188 characters including:

• Khmer Consonants: ក ខ គ ឃ ង ច ឆ ជ ឈ ញ ដ ឋ ឌ ឍ ណ ត ថ ទ ធ ន ប ផ ព ភ ម យ រ ល វ ស ហ ឡ អ
• Khmer Vowels: ា ិ ី ឹ ឺ ុ ូ ួ ើ ឿ ៀ េ ែ ៃ ោ ៅ ំ ះ ៈ
• Khmer Numerals: ០ ១ ២ ៣ ៤ ៥ ៦ ៧ ៨ ៩
• Latin Characters: A-Z, a-z, 0-9
• Punctuation: . , ! ? - ( ) [ ] « » ™ ® etc.
• Khmer Symbols: ។ ៕ ៖ ៗ ៉ ៊ ់ ៌ ៍ ៏ ័ ្

🚀 Quick Start

Installation

pip install paddlepaddle paddleocr opencv-python

Basic Usage

from paddleocr import PaddleOCR
import cv2

# Initialize OCR with custom Khmer model
ocr = PaddleOCR(
    use_angle_cls=True,
    lang='ch',  # Use Chinese as base language
    rec_model_dir='path/to/model',  # Directory containing inference files
    rec_char_dict_path='khmer_char_dict.txt',
    show_log=False
)

# Process image
result = ocr.ocr('khmer_text_image.jpg', cls=True)

# Extract results
for idx in range(len(result)):
    res = result[idx]
    if res is None:
        continue
    for line in res:
        text = line[1][0]  # Recognized text
        confidence = line[1][1]  # Confidence score
        print(f'Text: {text}, Confidence: {confidence:.3f}')

Command Line Usage

# Download model files to a directory
# Then use PaddleOCR tools:

python tools/infer/predict_rec.py \
    --image_dir="your_khmer_image.png" \
    --rec_model_dir="path/to/model" \
    --rec_char_dict_path="khmer_char_dict.txt"

📁 Files Included

File	Size	Description
inference.pdiparams	~106MB	Main model weights
inference.yml	~2KB	Model configuration
inference.json	~1KB	Model metadata
khmer_char_dict.txt	~2KB	Character dictionary (188 characters)
training_config.yml	~2KB	Original training configuration

🔧 Training Details

Dataset Characteristics

• Text Length: 3-5 words per image (optimized for short segments)
• Image Size: 600×80 pixels (training), resized to 320×32 for inference
• Font: KhmerOS TTF
• Background: White background with black text
• Augmentation: Clean, blurred, noisy, and noise+blur variants

Training Configuration

• Epochs: 30 (best model at epoch 29)
• Optimizer: Adam with β₁=0.9, β₂=0.999
• Learning Rate: Cosine scheduling (initial: 0.001)
• Batch Size: 32
• Loss Function: CTC Loss
• Regularization: L2 (factor: 4e-05)

💡 Usage Tips

Best Practices

1. Image Quality: Use high-contrast images with clear text
2. Text Length: Optimal for 3-5 word segments (model's training focus)
3. Resolution: Images should be reasonably sized (not too small)
4. Preprocessing: Consider using text detection for full documents

For Long Text Documents

Since this model is optimized for short segments, for full documents:

1. Use Text Detection: Combine with PaddleOCR's detection model
2. Segment Text: Break long lines into 3-5 word chunks
3. Post-process: Combine results from multiple segments

# Example for full document processing
ocr = PaddleOCR(
    use_angle_cls=True,
    lang='ch',
    det_model_dir='path/to/detection/model',  # Add detection model
    rec_model_dir='path/to/this/model',       # This Khmer recognition model
    rec_char_dict_path='khmer_char_dict.txt'
)

# This will detect text regions AND recognize them
result = ocr.ocr('full_document.jpg', cls=True)

🔄 Model Conversion

This model was exported from PaddlePaddle training format to inference format:

# Original export command used:
python tools/export_model.py \
    -c pretrainoutput/config.yml \
    -o Global.pretrained_model=pretrainoutput/best_accuracy.pdparams \
    Global.save_inference_dir=pretrainoutput/inference

🛠️ Requirements

paddlepaddle>=2.4.0
opencv-python>=4.5.0
numpy>=1.19.0
pillow>=8.0.0

@misc{khmer-ocr-2025,
  title={Khmer OCR Recognition Model},
  author={[Your Name]},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/[your-username]/khmer-ocr}}
}