hazemmohamedsayed123/OCR-table-extractor

OCR Table Extractor

Overview

This project extracts structured tables from scanned PDF documents and images.

PaddleOCR-VL is used as the primary table extraction model. Gemini independently analyzes the same document and is used as a verification model to identify cells that may require manual review.

Features

• Scanned PDF and image upload
• Automatic table extraction
• English and Arabic table support
• Excel workbook generation
• Highlighted disagreements between models
• Review Needed worksheet
• Summary worksheet
• Supabase storage and job tracking

Base Model

This project uses the official PaddleOCR-VL model.

The PaddleOCR-VL weights were not trained or created by this project. They are downloaded automatically through PaddleOCR when the backend starts.

No custom fine-tuned checkpoint has been uploaded yet.

Processing Pipeline

1. A user uploads a PDF or image.
2. PaddleOCR-VL detects and extracts the tables.
3. Gemini extracts the same document independently.
4. The two outputs are compared.
5. PaddleOCR remains the primary output.
6. Different cells are highlighted for manual review.
7. A reviewed Excel workbook is generated.

Supported Inputs

• PDF
• PNG
• JPG
• JPEG
• WebP

Supported Languages

• English
• Arabic
• Mixed English and Arabic tables

Current Limitations

• Arabic OCR may contain spelling or character errors.
• Complex merged cells may be interpreted differently.
• Tables continuing across pages may require post-processing.
• Gemini availability depends on its external API.
• The public Vercel deployment currently contains the frontend only.
• The PaddleOCR backend requires a separate machine with suitable processing capacity.

Source Code

https://github.com/hazemmohamedsayed/OCR-table-extractor

Live Frontend

https://ocr-table-extractor-lemon.vercel.app

Future Work

• Fine-tune a model using company-specific bilingual forms.
• Improve Arabic text normalization.
• Improve multi-page table merging.
• Deploy the FastAPI OCR backend publicly.