SmartKhOCR 🇰🇭

SmartKhOCR is a state-of-the-art Khmer Optical Character Recognition (OCR) and Language Correction engine built with Deep Learning Vision Transformers (ViT) and Khmer phonetic language shaping models.

It provides real-time line-by-line streaming, native PDF multi-page recognition, and contextual Coeng (ជើង) & compound auto-correction.

✨ Features

• 🚀 Deep Learning Vision Engine: Vision Transformer + CTC/Beam Search specifically fine-tuned on modern document scripts, headers, body text, and complex font styles.
• ✨ Coeng & Compound Auto-Correction: Context-aware corrector that fixes hanging Coeng (e.g., ឆ្ាំ $\to$ ឆ្នាំ), accidental subscripts (សាល្ា $\to$ សាលា), and missing Coeng compounds (រាជធានីភំពេញ $\to$ រាជធានីភ្នំពេញ).
• ⚡ Ultra-Fast GPU Streaming: Real-time line-by-line and page-by-page streaming generator (13.9 ms/line CTC decoding on NVIDIA RTX GPUs).
• 📄 Native PDF & Image Pipeline: Direct multi-page PDF processing with on-the-fly rendering and lightweight RAM footprint.
• 🔍 Side-by-Side Comparison Studio: Integrated web dashboard with bounding box visualization and parallel Tesseract OCR (khm+eng) benchmark comparisons.
• 💻 CLI & REST API Ready: Built-in CLI tool and FastAPI Server-Sent Events (SSE) streaming server.

📦 Installation

# Clone the repository
git clone https://github.com/your-username/SmartKhOCR.git
cd SmartKhOCR

# Install dependencies using uv or pip
uv pip install -e .
# or
pip install -e .

🚀 Quickstart (Python API)

1. Basic Image OCR

from SmartKhOCR import SmartKhOCR

# Initialize OCR engine (auto-selects CUDA if available)
model = SmartKhOCR(decode_method="fast")

# Recognize text with Coeng auto-correction
result = model("chapter.png", correct_word=True)

print("Recognized Text:")
print(result.text)

# Save annotated visualization with detected line boxes
result.save("output_annotated.jpg")

2. Multi-Page PDF Processing

from SmartKhOCR import SmartKhOCR

model = SmartKhOCR()

# Process multi-page PDF directly
pages = model("document.pdf", correct_word=True)

for idx, page in enumerate(pages, 1):
    print(f"\n--- Page {idx} ---")
    print(page.text)

3. Real-Time Streaming Generator

from SmartKhOCR import SmartKhOCR

model = SmartKhOCR(decode_method="fast")

# Stream lines in real-time as they are decoded
for line in model("document.png", stream=True, correct_word=True):
    print(f"[{line.box}] {line.text}")

🖥️ Command Line Interface (CLI)

SmartKhOCR installs a command-line tool directly into your environment:

# Run OCR on an image or PDF
SmartKhOCR document.png --correct-word --output-txt transcript.txt --save boxes.jpg

# Export model to high-performance inference engines
SmartKhOCR --export engine       # TensorRT engine
SmartKhOCR --export onnx         # ONNX graph
SmartKhOCR --export torchscript  # TorchScript JIT

🌐 Web Studio & Streaming REST API

SmartKhOCR comes with a real-time web dashboard and SSE streaming API:

uv run fastapi run app_ui.py --host 0.0.0.0 --port 8000

1. Web Dashboard: Open http://localhost:8000 in your browser.
2. REST API Streaming:

curl -N -X POST http://localhost:8000/api/ocr/stream \
  -F "file=@document.pdf" \
  -F "correct_word=true" \
  -F "fast_mode=true" \
  -F "compare_tesseract=true"

🏷️ Package Architecture

You can import SmartKhOCR using either convention:

# Direct class import (Recommended)
from SmartKhOCR import SmartKhOCR

# Snake_case module import
from smart_kh_ocr import SmartKhOCR

# Backward-compatibility alias
from khmer_ocr import KhmerOCR

📄 License

MIT License. Designed and developed for high-accuracy Khmer text processing.