Arabic Document OCR Project

Local and private OCR system for extracting text from Arabic documents using local models. All processing runs 100% locally — no data is ever sent to any external server.

---

Environment

• Python: 3.11
• Package Manager: Conda
• Environment Name: test_ocr
• OS: Windows
• GPU: GeForce MX350 (2GB VRAM, CUDA 10.2)
• Colab: Google Colab T4 (15GB VRAM) for large models

---

Project Structure

testing_ocr/
├── main.py                      # CLI runner for OCR experiments
├── handwriting_pipeline.py      # Handwriting detection pipeline
├── test_easyocr.py             # EasyOCR basic test
├── test_easyocr_enhanced.py    # EasyOCR with preprocessing comparison
├── test_lightonocr.py          # LightOnOCR test
├── test_qari.py                # Qari-OCR test (Colab)
├── README.md
├── requirements.txt
├── config/
│   ├── __init__.py
│   └── settings.py
├── src/
│   ├── __init__.py
│   ├── ocr_engine.py           # Unified OCR engine (4 models)
│   ├── preprocessor.py         # Image preprocessing (7 methods)
│   ├── handwriting_detector.py # YOLOv8 handwriting detection
│   └── wakala_parser.py
├── files/                      # Test documents (not on GitHub)
├── output/                     # Results (not on GitHub)
└── tests/

---

Hardware Constraints

• GPU: MX350 with 2GB VRAM — too small for most AI models
• Models must run on CPU locally (slow) or on Google Colab T4 (free)
• 4-bit quantization required for 2B+ models on Colab

---

Models Tested

Experiment Results Summary

---

Detailed Experiment Results

Experiment 1A: LightOnOCR-2-1B (Original)

• Date: 2026-02-15
• File: وكالة الكترونية 009.pdf
• Preprocessing: None
• Device: CPU (local)
• Load time: 11.9s | OCR time: 845.6s (~14 min)
• Output: 2433 chars
• Strengths: Excellent structure detection (tables, sections, HTML output)
• Weaknesses: Arabic words weak (الموثقين→المونتين, مقفلة→مقللة)
• Numbers: Partially wrong (missing digits)
• Score: 6/10

Experiment 1B: LightOnOCR-2-1B (Denoise)

• Date: 2026-02-15
• Same file, added denoise + CLAHE preprocessing
• OCR time: 966.7s (slower)
• Output: 2448 chars
• Result: Some words improved, others regressed
• Conclusion: Preprocessing does NOT help on clean electronic documents
• Score: 5.5/10

Experiment 2A: Qari-OCR v0.3

• Date: 2026-02-16
• Model: NAMAA-Space/Qari-OCR-v0.3-VL-2B-Instruct
• Device: Colab T4 GPU, 4-bit quantization, 150 DPI
• OCR time: 58.1s (14x faster than LightOnOCR)
• Output: 1288 chars
• Strengths: Better Arabic (مقفلة✅, المتاحة✅, إلغاء✅)
• Weaknesses: Missing critical data (names, IDs, dates, tables)
• Score: 5/10

Experiment 7: EasyOCR

• Date: 2026-02-16
• Device: CPU local (no GPU)
• Load time: 4.9s

7A: Electronic Document (وكالة الكترونية)

• OCR time: 102.4s (DPI 300) / 86.1s (DPI 400)
• Detections: 68-72
• Avg Confidence: 75.0%
• Strengths: Arabic words good, full paragraphs readable
• Weaknesses: Dates/numbers broken, edge words truncated

7B: EasyOCR Preprocessing Comparison (DPI 400)

Best method: full (denoise → CLAHE → sharpen → Otsu threshold)

7C: Handwritten Document (عقد ايجار ورقي)

• OCR time: 114.7s
• Detections: 77
• Avg Confidence: 15.6%
• Result: FAILED — almost all text unreadable
• Score: 1.5/10

---

Handwriting Detection Experiments

YOLOv8n Handwriting Detection

• Model: armvectores/yolov8n_handwritten_text_detection
• Device: CPU local
• Result on صك ورقي: 97 regions at 25% confidence → 8 regions at 60%
• Correctly found: Signatures and stamps at bottom
• Missed: Handwritten text inside form fields
• Conclusion: Trained on English docs, not suitable for Arabic mixed documents

Computer Vision Approaches (Colab)

• Color separation: Failed — printed and handwritten text same color
• Stroke width analysis: Partially worked — found signatures/stamps only
• Conclusion: CV approaches not sufficient for mixed Arabic documents

---

Key Findings

1. Electronic documents: EasyOCR with full preprocessing gives best results (76.1% confidence)
2. Handwritten documents: No model tested so far works well (best: 15.6%)
3. Preprocessing helps slightly on electronic docs but NOT significantly
4. DPI 400 better than 300 for EasyOCR
5. Arabic handwriting is extremely challenging for all tested models
6. VLM models (2B+) tend to hallucinate or loop on Arabic documents
7. Detection-based approaches (YOLO) miss handwriting inside form fields

---

---

Next Steps

1. Test Arabic-English-handwritten-OCR-v3 on Colab with proper settings
2. Test HATFormer (small, handwriting specialist)
3. Test Surya OCR locally
4. Consider fine-tuning Qari-OCR on Saudi legal documents
5. Build combined pipeline: detection → separation → OCR

---

Setup

# Create environment
conda create -n test_ocr python=3.11
conda activate test_ocr

# Install dependencies
pip install easyocr
pip install torch torchvision
pip install git+https://github.com/huggingface/transformers
pip install pypdfium2 Pillow opencv-python numpy
pip install ultralytics huggingface_hub

# Run EasyOCR test
python test_easyocr_enhanced.py

# Run handwriting detection
python handwriting_pipeline.py --file "files/doc.pdf" --confidence 0.6 --save

---

Notes

• All files in files/ are confidential — never upload to Git
• All processing must be 100% local (except Colab for large models)
• No cloud APIs (OpenAI, Google Vision, etc.)
• Output files shared via OneDrive (not on GitHub)
• GitHub repo: https://github.com/Meshal-Css/OCR_task.git