OCR Benchmarking System for Scientific Literature

This repository contains a comprehensive benchmarking framework that evaluates and compares OCR systems on scientific literature. The project successfully benchmarks 3 OCR systems (Marker, Docling, PyMuPDF) with Marker achieving 82.6% accuracy as the best performer.

Deliverables Completed:

• ✅ 3 OCR systems benchmarked (Marker + 2 others as requested)
• ✅ Google Colab Notebook with findings and analysis
• ✅ Comprehensive metrics for scientific document processing
• ✅ Processing time analysis and performance comparison
• ✅ Real scientific literature used as test dataset

The benchmarking system evaluates OCR systems across multiple dimensions:

• Text Extraction Accuracy: Character, word, and line-level accuracy
• Processing Performance: Speed, efficiency, resource usage
• Scientific Content Handling: Formulas, notation, references, citations

• Document Elements: Title, authors, abstract detection
• Section Organization: Headers, paragraphs, reading order
• Scientific Elements: Equations, tables, figures, references
• Layout Preservation: Document hierarchy and structure

• Marker ⭐ - Priority system (82.6% accuracy, 39s avg processing)
• Docling - IBM's document AI (81.3% accuracy, 91s avg processing)
• PyMuPDF - Direct text extraction baseline (<1s processing)

Result: Marker demonstrates the best overall performance for scientific literature OCR, balancing high accuracy with reasonable processing speed.

papers-ocr-benchmarks/
├── 📓 OCR_Benchmark_Scientific_Literature.ipynb  # 🎯 Google Colab Notebook (MAIN DELIVERABLE)
├── 📄 README.md                                  # Project documentation
├── 📋 requirements.txt                           # Python dependencies
│
├── 📁 scripts/                                   # Benchmark scripts
│   ├── ocr_benchmark_gpu_optimized.py           # Main benchmark script
│   └── setup_gpu_environment.py                 # Environment setup
│
├── 📁 pdfs/                                      # Test dataset (3 scientific papers)
│   ├── 2014-Combining_organophosphate_treated_wall_linings...pdf
│   ├── Allossogbe_et_al_2017_Mal_J.pdf
│   └── Somboon_et_al_1995_Trans_RSTMH.pdf
│
├── 📁 results/                                   # Benchmark results
│   ├── benchmark_results.csv                    # Detailed metrics
│   ├── benchmark_summary.csv                    # Summary statistics
│   ├── benchmark_visualization.png              # Performance charts
│   └── [individual_ocr_outputs].txt             # Raw OCR extractions
│
├── 📁 docs/                                      # Project documentation
│   └── Enhanced AI OCR Extraction Pipeline for Scientific Literature.md
│
└── 📁 output/                                    # Generated results directory

1. View Results: Open OCR_Benchmark_Scientific_Literature.ipynb in Google Colab
2. Check Metrics: Review results/benchmark_summary.csv
3. Inspect Outputs: Examine individual OCR outputs in results/

# 1. Setup environment
python scripts/setup_gpu_environment.py

# 2. Run benchmark
python scripts/ocr_benchmark_gpu_optimized.py

# 3. Analyze document structure (NEW)
python scripts/structure_parser.py

# 4. Check results in results/ and examples/outputs/ directories

• OCR_Benchmark_Scientific_Literature.ipynb - Google Colab notebook with complete analysis
• results/latest_benchmark_results.csv - Latest benchmark results summary
• examples/outputs/structure_comparison.csv - Structure parsing analysis
• STRUCTURE_ANALYSIS_REPORT.md - Detailed structure parsing evaluation

• scripts/ocr_benchmark_gpu_optimized.py - Main benchmark script with GPU optimization
• scripts/structure_parser.py - Document structure analysis tool
• scripts/setup_gpu_environment.py - Environment setup and dependency checking

• pdfs/ - Test dataset of 3 scientific papers
• results/ - Complete benchmark results and individual OCR outputs
• examples/outputs/ - Structured JSON outputs for each OCR system

• Best Accuracy: Marker (82.6% character accuracy)
• Fastest Processing: PyMuPDF (<1s per document)
• Best Balance: Marker (high accuracy + reasonable speed)

• Citations Detected: All systems successfully identify reference citations
• Figures/Tables: Good preservation of figure and table references
• Mathematical Content: Basic formula detection implemented

• Marker: 39s average (best AI-based performance)
• Docling: 91s average (thorough but slower)
• PyMuPDF: <1s (direct text extraction baseline)

pip install -r requirements.txt

Key packages:

• marker-pdf - Marker OCR system
• docling - IBM Docling system
• PyMuPDF - PDF text extraction
• pandas, numpy - Data analysis
• torch - GPU acceleration (optional)

• Python 3.8+
• 8GB+ RAM recommended
• GPU optional (automatic CPU fallback)
• 5GB+ storage for results