deepdoctection/deepdoctection

What it solves

Deepdoctection is designed to handle the complex task of document understanding. It solves the problem of converting unstructured scan or PDF documents into structured data by orchestrating various AI models for layout analysis, OCR, and classification.

How it works

The library acts as an orchestration layer that builds pipelines for document extraction. It integrates multiple specialized tools and models:

• Layout Analysis & Table Recognition: Uses PyTorch with Detectron2 and Transformers.
• OCR: Supports Tesseract, DocTr, and AWS Textract.
• ** uma Document & Token Classification:** Employs the LayoutLM family, LiLT, and Bert-style models.
• Text Mining: Uses pdfplumber for native PDFs.
• Preprocessing: Includes deskewing and rotating images via jdeskew or Tesseract.
• Language Detection: Uses transformer-based models for language identification.

Who it’s for

Developers and data scientists who need to build automated document processing workflows, extract structured information from PDFs and scans, or fine-tune pre-trained document AI models for specific tasks.

Highlights

• Modular Pipeline: Orchestrates multiple OCR and layout analysis tools into a single workflow.
• Broad Model Support: Integrates with Hugging Face Hub for models like LayoutLM, BERT, and RoBERTa.
• PyTorch-only: Recent refactoring ensures consistent support across deep learning models.
• Fine-tuning Capabilities: Allows users to fine-tune object detection and classification models and evaluate the resulting pipelines.