kiri-ocr 0.1.4

A lightweight OCR library for Khmer and English documents

Kiri OCR 📄

Kiri OCR is a lightweight, OCR library for English and Khmer documents. It provides document-level text detection, recognition, and rendering capabilities in a compact package (~13MB model).

✨ Key Features

• Lightweight: Only ~13MB model size (Lite version).
• Bi-lingual: Native support for English and Khmer (and mixed).
• Document Processing: Automatic text line and word detection.
• Robust Detection: Works on both light and dark backgrounds (Dark Mode support).
• Easy to Use: Simple Python API.
• Visualizations: Generate annotated images and HTML reports.

📊 Dataset

The model is trained on the mrrtmob/km_en_image_line dataset, which contains 5 million synthetic images of Khmer and English text lines.

📈 Benchmark

Results on synthetic test images (10 popular fonts):

📦 Installation

Install easily via pip:

pip install kiri-ocr

Or install from source:

git clone https://github.com/mrrtmob/kiri-ocr.git
cd kiri-ocr
pip install .

💻 Usage

CLI Tool (Inference)

Run OCR on an image and save results:

kiri-ocr predict path/to/document.jpg --output results/

(Or simply kiri-ocr path/to/document.jpg)

Python API

from kiri_ocr import OCR

# Initialize (loads from Hugging Face automatically)
ocr = OCR()

# Extract text
text, results = ocr.extract_text('document.jpg')
print(text)

🎓 Training a New Model

Follow this guide to train a custom model from scratch.

Step 1: Generate Training Data

Create synthetic training images from a text file.

1. Prepare text file: Create data/textlines.txt with your training text (one sentence per line).

2. Generate dataset:

   kiri-ocr generate \
       --train-file data/textlines.txt \
       --output data \
       --fonts-dir fonts \
       --augment 1 \
       --random-augment
   

   • --fonts-dir: Directory containing .ttf files (Khmer/English fonts).
   • --augment: How many variations to generate per line (e.g., 2).
   • --random-augment: Apply random noise/rotation even if augment is 1.

Custom Dataset Structure

If you have your own data (not generated), organize it as follows:

data/
  ├── train/
  │   ├── labels.txt       # Tab-separated: filename <tab> text
  │   └── images/          # Image files
  │       ├── img_001.png
  │       ├── img_002.jpg
  │       └── ...
  └── val/
      ├── labels.txt
      └── images/

Format of labels.txt:

img_001.png    Hello World
img_002.jpg    This is a test

Note: Images must be in an images/ subdirectory relative to the labels.txt file.

Step 2: Train the Model

You can train using CLI arguments or a configuration file.

Option A: Using Configuration File (Recommended)

1. Generate default config:

   kiri-ocr init-config -o config.json
   

2. Edit config.json to adjust hyperparameters (epochs, batch size, etc.).
3. Start training:

   kiri-ocr train --config config.json
   

Option B: Using CLI Arguments

kiri-ocr train \
    --train-labels data/train/labels.txt \
    --val-labels data/val/labels.txt \
    --epochs 100 \
    --batch-size 32 \
    --device cuda

Option C: Training with Hugging Face Dataset

You can train directly using a dataset from Hugging Face Hub. The dataset should contain image and text columns.

kiri-ocr train \
    --hf-dataset mrrtmob/km_en_image_line \
    --epochs 50 \
    --batch-size 32

Advanced HF Options:

• --hf-train-split: Specify training split name (default: "train").
• --hf-val-split: Specify validation split name. If not provided, it tries "validation", "val", "test", or automatically splits the training set.
• --hf-val-percent: Percentage of training data to use for validation if no validation split is found (default: 0.1 for 10%).
• --hf-image-col: Column name for images (default: "image").
• --hf-text-col: Column name for text labels (default: "text").
• --hf-subset: Dataset configuration/subset name (optional).

To use a specific subset/config (if the dataset has multiple):

kiri-ocr train \
    --hf-dataset mrrtmob/km_en_image_line \
    --hf-subset default \
    ...

Fine-Tuning

To fine-tune an existing model on new data:

kiri-ocr train \
    --config config.yaml \
    --from-model models/model.kiri

This loads the weights from models/model.kiri before starting training. Useful for domain adaptation or adding languages.

The trained model will be saved to models/model.kiri (or specified output_dir).

☕ Support

If you find this project useful, you can support me here:

• Buy Me a Coffee
• ABA Payway

⚖️ License

Apache License 2.0.