🔬 Doctane by Zeross

Multimodal Intelligent Document Analysis and Understanding System

---

Doctane is a vision-language-based pipeline built to automate document analysis, including OCR, layout understanding, semantic extraction, and captioning — all under one modular framework. Designed for business documents, forms, invoices, and scanned reports.

---

✨ Key Features

• End-to-End OCR: Unified pipeline for text detection and recognition
• Robust Detection: Handles both straight and rotated text using segmentation-based models
• Accurate Recognition: High-accuracy text recognition with support for custom vocabularies
• Synthetic Data Generation: Includes WordGenerator to create training data on-the-fly
• Layout-Aware: Automatically detects page orientation and can straighten pages before analysis
• Structured Output: Exports results as rich, hierarchical Document objects (Page, Block, Line, Word) and supports hOCR format
• Distributed Training: Comes with DDP (Distributed Data Parallel) scripts for faster training on multiple GPUs
• Extensible: Modular design with a DocumentBuilder and hooks for custom post-processing
• Multi-lingual Support: Features language detection and allows training with different vocabularies

---

🧬 Processing Pipeline

The pipeline processes documents through the following stages:

1. Preprocessing: Handles page orientation detection and optional straightening
2. Text Detection: A segmentation-based model (e.g., LinkNet) identifies word or line bounding boxes (both straight and rotated)
3. Text Recognition: Crops from the detection stage are fed into a recognition model to transcribe the text
4. Document Assembly: The final output is a structured Document object containing pages, blocks, lines, and words with their geometry and confidence scores. The system also supports Key Information Extraction (KIE)

---

🏗️ Supported Models

Detection Models

Model	Description
LinkNet	Fast and lightweight segmentation
DeepLabV3+	High accuracy for complex layouts
SegFormer	State-of-the-art transformer-based
Faster R-CNN	Region-based detection

Recognition Models

Model	Description
SAR	Sequence Approximation Recognition
ViTSTR	Vision Transformer for STR
CRNN	CNN + RNN + CTC
TrBA	Transformer-based Recognition
MASTER	Multi-Aspect Self-Attention
ABINet	Attention-based Iterative Network
VisionLAN	Vision-Language Awareness Network
DiT	Diffusion Transformer

---

🚀 Quick Start

One-Click Setup (Windows)

git clone https://github.com/Purushothaman-natarajan/doctane.git
cd doctane
run.bat

Manual Setup

# Clone the repository
git clone https://github.com/Purushothaman-natarajan/doctane.git
cd doctane

# Create virtual environment
python -m venv venv
source venv/bin/activate  # Linux/Mac
venv\Scripts\activate      # Windows

# Install dependencies
pip install -r requirements.txt
pip install -e .

# Run the web interface
streamlit run app.py

Open http://localhost:8501 in your browser to use Doctane.

---

📖 Documentation

• Documentation - Comprehensive API and usage guide
• Architecture - System design and component details
• Developer Guide - Setup, training, and contribution guide
• Profile - Developer information

---

📂 Project Structure

Doctane/
├── index.html              # Landing page
├── docs.html               # Documentation
├── architecture.html       # System architecture
├── developer-guide.html    # Developer guide
├── profile.html            # Developer profile
├── app.py                  # Streamlit web interface
├── run.bat                 # Windows launcher
│
├── doctane/                # Main package
│   ├── predictor/          # Predictor classes
│   ├── ocr_pipeline/      # OCR pipeline
│   ├── models/             # Model definitions
│   │   ├── detection/      # Detection models
│   │   ├── recognition/    # Recognition models
│   │   └── classification/ # Classification models
│   ├── datasets/           # Dataset loaders
│   ├── train/              # Training scripts
│   ├── utils/              # Utilities
│   └── evaluate/           # Evaluation
│
├── configs/                 # YAML configurations
├── data/                    # Data directory
├── scripts/                 # Utility scripts
├── tests/                   # Unit tests
├── notebooks/               # Jupyter notebooks
└── requirements.txt         # Dependencies

---

💻 Usage Examples

Python API

import numpy as np
from PIL import Image
from doctane.ocr_pipeline.ocr_predictor import OCRPredictor

# Load image
image = Image.open("document.jpg").convert("RGB")
np_image = np.array(image)

# Run OCR
predictor = OCRPredictor()
output = predictor([np_image])

# Access results
for page in output.pages:
    print(page)

Command Line

python scripts/run_inference.py \
    --input_dir data/raw \
    --output_dir results/outputs \
    --model_type seg_linknet_resnet50

---

📊 Supported Datasets

• Custom Datasets: Easily train on your own data using the DetectionDataset and RecognitionDataset classes
• Synthetic Data: The WordGenerator can create arbitrary amounts of training data
• Public Datasets (Planned): FUNSD, PubLayNet, DocVQA

---

🛠️ Training

Text Detection

python train/text_detection/train_detection.py \
    --config configs/detection.yaml \
    --epochs 100

Text Recognition

python train/text_recognition/train_recognition.py \
    --config configs/recognition.yaml \
    --epochs 50

Distributed Training (DDP)

torchrun --nproc_per_node=4 \
    train/text_detection/train_detection_ddp.py \
    --config configs/detection.yaml

---

🔧 Configuration

Key environment variables:

Variable	Description	Default
BUILD_VERSION	Package version	0.1.0a0
CHECKPOINT_DIR	Model checkpoints	./checkpoints
DATA_DIR	Training data	./data
LOG_DIR	Training logs	./logs

---

🤝 Contributing

1. Fork the repository
2. Create a feature branch (git checkout -b feature/your-feature)
3. Make changes and add tests
4. Run linting and type checking
5. Submit a pull request

---

📈 Future Roadmap

• End-to-end document captioning
• Visual layout annotation
• Web Dashboard for interaction
• Multi-language support expansion
• PDF export functionality

---

🙏 Acknowledgments

• Hugging Face LayoutLM, Donut, doctr
• Tesseract OCR, EasyOCR
• pytorch-segmentation-models
• PyMuPDF, PDFPlumber

---

📜 License

Apache License 2.0 - See LICENSE for details.

---

👨‍💻 Developer

Purushothaman Natarajan

• GitHub: @Purushothaman-natarajan
• Portfolio: purushothaman-natarajan.github.io
• Email: purushothamanprt@gmail.com

---

Made with ❤️ by Doctane