OCR Comparison: Tesseract vs Amazon Textract

This repository contains scripts, configuration, and sample results from testing OCR quality across a variety of document types in the New York Philharmonic Archives.

Objective

To evaluate whether we can rely primarily on open-source Tesseract or whether Amazon Textract is needed for complex documents (e.g., overlaid stamps, handwriting).

Summary of Findings

• ✅ Tesseract (with tuned settings) performs well on clean, typed materials.
• ❌ Tesseract struggles with:
  • Overlaid stamps (e.g., "COPY")
  • Handwriting
• ✅ Textract produced much better results in those cases.

Recommended Strategy

• Use Tesseract for most documents with:

  • --psm 3 --oem 1
  • Grayscale JPEGs converted from JP2 with enhanced contrast

• Use Amazon Textract for:

  • Stamped letters
  • Handwritten annotations

Contents

• ocr_comparison.py: main OCR test harness
• test_images/: location for input images (JP2s, JPGs, etc.)
• results/: sample outputs for Tesseract and Textract

How to Run

1. Install Python dependency:

pip install boto3

2. Configure your AWS credentials (for Textract):

aws configure --profile textract-test

3. Run the script:

python ocr_comparison.py