PDF Table Extraction with OCR

A tool for extracting tables from PDF documents using image-based techniques and OCR.

Features

• Loads PDF files and converts them to images
• Detects tables in PDF pages using computer vision
• Crops tables from pages with configurable padding
• Identifies rows and columns using line detection
• Extracts text using PaddleOCR with multi-language support
• Exports table data to CSV files
• Merges tables with identical structures into consolidated CSVs
• Progress tracking with visual feedback
• Comprehensive logging of operations

Key Features

Intelligent CSV Merging

• Tables are grouped by column structure
• Multiple merged files created for different table formats
• Header row deduplication during merging
• Preserves original table order within structure groups

Logging

• Detailed operations log in table_extractor.log
• Includes both application and OCR engine logs
• Overwritten for each new execution

Progress Tracking

• Real-time progress bars for:
  • PDF loading
  • Table detection
  • OCR processing
  • File saving
• Clean completion message with timestamp

Output Notes

• Tables with different column counts are saved in separate merged files
• First table's header is used as reference for merging
• Subsequent tables with matching headers are merged without header row
• Tables with unique headers are preserved as-is
• Empty rows/columns are automatically trimmed from CSV outputs

Installation

1. Clone the repository
2. Install the required dependencies:

pip install -r requirements.txt

or Using Astral UV

uv sync

Note: you'll need to install Astral UV first if not already installed.

Note for Windows users

For Windows, you might need to install Poppler separately:

1. Download the Poppler binary from: https://github.com/oschwartz10612/poppler-windows/releases/
2. Extract it to a folder (e.g., C:\poppler)
3. Add the bin directory to your PATH: C:\poppler\bin

Usage

Run the table extraction tool with:

python main.py path_to_your_pdf.pdf

Optional arguments

• --output_dir, -o: Output directory (default: 'data/output')
• --dpi: DPI for PDF to image conversion (default: 300)
• --denoise/--no-denoise: Toggle image denoising (default: True)
• --morph-close/--no-morph-close: Toggle morphological closing (default: True)
• --crop-padding: Padding around cropped tables in pixels (default: 1)
• --lang: OCR language code (default: 'en')
• --use-gpu: Use GPU acceleration for OCR (requires compatible hardware)

Example:

python main.py financial_report.pdf --output_dir results --dpi 400 --lang en --use-gpu

Output Structure

The tool creates the following directory structure:

output
├── merged_table.csv      # All tables merged into single CSV
├── cropped/              # Cropped table images
│   ├── page_0_table_0.png
│   ├── page_0_table_1.png
│   └── ...
├── csv/                  # Individual table CSV files
│   ├── page_0_table_0.csv
│   ├── page_0_table_1.csv
│   └── ...
├── detected/             # Pages with table detection boxes
│   ├── page_0.png
│   ├── page_1.png
│   └── ...
├── original/             # Original page images
│   ├── page_0.png
│   ├── page_1.png
│   └── ...
└── structure/            # Tables with grid structure
    ├── page_0_table_0.png
    ├── page_0_table_1.png
    └── ...

Each subdirectory contains specific information about the processing pipeline:

• original/: Original PDF pages as images
• detected/: Pages with green boxes showing detected table regions
• cropped/: Individual cropped tables from each page
• structure/: Tables with detected grid structure (red lines)
• csv/: CSV files containing the extracted text from each table
• merged_table.csv: All tables combined into a single CSV file, ordered by page and table number

Acknowledgement

• PaddleOCR for optical character recognition.