Data Processing with OCR

Overview

This project provides multiple approaches for extracting text from PDF documents using both structured parsing and OCR techniques. It is designed for document preprocessing workflows such as Retrieval-Augmented Generation (RAG), data pipelines, and information extraction systems.

The project includes three main scripts:

• docling.py – Structured PDF parsing using Docling
• tesseract.py – OCR-based extraction using Tesseract
• easyocr.py – OCR-based extraction using EasyOCR

Each script processes PDFs and outputs text in Markdown and/or JSON formats.

---

Project Structure

project_root/

│

├── input_folder/ # Input PDFs for all scripts

│

├── output_folder/ # Rename or organize outputs here

│

├── docling.py

├── tesseract.py

├── easyocr.py

└── README.md

---

Important Setup Note

Before running the scripts, ensure:

• All input directories are renamed to: input_folder
• All output directories are renamed to: output_folder

If your scripts still reference older folder names, update them accordingly or rename your folders to match.

---

1. Docling Pipeline (docling.py)

Description

Uses Docling's document parser to extract structured text directly from PDFs without OCR.

Features

• Preserves document structure
• Outputs clean Markdown
• Fast for machine-readable PDFs

Input

• Folder: input_folder/

Output

• Markdown files saved to your configured output_folder/

Workflow

1. Load PDFs from input folder
2. Convert using Docling
3. Export structured text to Markdown

---

2. Tesseract OCR Pipeline (tesseract.py)

Description

Uses Tesseract OCR to extract text from scanned PDFs by converting pages into images.

Features

• Image preprocessing using OpenCV
• Optimized for structured forms
• Outputs both Markdown and JSON
• Page-level text extraction

Requirements

• Tesseract OCR installed
• Path configured: pytesseract.pytesseract.tesseract_cmd = "C:\Program Files\Tesseract-OCR\tesseract.exe"

Input

• Folder: input_folder/

Output

• Markdown and JSON saved to your configured output_folder/

Workflow

1. Convert PDF pages to images (300 DPI)
2. Preprocess images (grayscale + thresholding)
3. Extract text using Tesseract
4. Clean and structure output
5. Save as Markdown and JSON

---

3. EasyOCR Pipeline (easyocr.py)

Description

Uses EasyOCR for deep learning-based text extraction from PDFs.

Features

• Confidence-based filtering
• Outputs structured JSON with confidence scores
• Good for complex or low-quality scans

Requirements

• Poppler installed (for PDF rendering)
• Path configured: POPPLER_PATH = "C:\Users..."

Input

• Folder: input_folder/

Output

• Markdown and JSON saved to your configured output_folder/

Workflow

1. Detect number of pages
2. Convert each page to image
3. Run EasyOCR text detection
4. Filter results by confidence threshold
5. Save results to Markdown and JSON

---

Comparison of Approaches

Method	Best For	Output Type	Speed	Accuracy
Docling	Digital PDFs (text-based)	Markdown	Fast	High
Tesseract	Scanned forms/documents	Markdown + JSON	Medium	Medium
EasyOCR	Complex or noisy images	Markdown + JSON	Slower	High

---

Use Cases

• Document ingestion for RAG systems
• Converting PDFs into structured datasets
• Extracting text from scanned forms
• Preprocessing for machine learning pipelines

---

Installation

Install required Python packages: pip install docling pytesseract pdf2image easyocr opencv-python pillow numpy

Additional System Dependencies

• Tesseract OCR (required for tesseract.py)
• Poppler (required for easyocr.py)

---

How to Run

Run Docling

python docling.py

Run Tesseract OCR

python tesseract.py

Run EasyOCR

python easyocr.py

---

Notes

• Ensure input_folder/ contains PDF files before running scripts
• Ensure output_folder/ exists or is correctly referenced in scripts
• OCR performance depends heavily on image quality and preprocessing

---

Future Improvements

• Add text chunking for RAG pipelines
• Integrate embeddings and vector databases
• Add support for batch processing with metadata tagging
• Improve layout-aware parsing