🛡️ OCR-Docker-Framework

High-Performance Optical Character Recognition & Performance Auditing Suite

An opinionated, production-ready OCR engine with automated performance validation.

📖 Overview

"In OCR, 'it reads' is not a valid test result. Accuracy is the only metric."

This framework automates the OCR validation process that professional applications require — turning raw pixel data into structured, audited text evidence. Built to be modular, testable, and deployable.

Sample Extraction

(The automated PaddleOCR engine rendering bounding boxes for auditing)

This framework is production-ready, optimized for containarized environments, ensuring that your data remains secure while providing high-accuracy text extraction.

The Problem it Solves

Many OCR implementations are "black boxes." This framework provides an audit trail for every extraction, allowing you to measure F1 Score and CER (Character Error Rate) against your own gold-standard datasets.

📂 Project Structure

ocr_docker/
├── .github/
│   └── workflows/      # Automated sync to Hugging Face Spaces
├── core/
│   ├── evaluators/     # F1 Score & CER character-level metrics
│   ├── ocr_engine.py   # High-level PaddleOCR orchestrator
│   └── utils.py        # Persistence & reporting utilities
├── data/
│   └── gold_standard/  # Reference images and ground-truth JSON
├── reports/            # Generated audit trails (JSON/Markdown)
├── scripts/
│   └── run_benchmark.py # CLI Entry Point for performance auditing
├── tests/              # Pytest unit testing suite
├── app.py              # Gradio web interface (HuggingFace Spaces)
├── requirements.txt    # Production dependencies
└── requirements-dev.txt # Test-only dependencies

🛠️ The Core Modules

1. Accuracy Evaluator (F1 Score)

Standard text extraction metrics focus on word overlap. This module ensures that every word in the source document is accurately represented, penalizing omissions and false positives.

2. Character Error Rate (CER)

For high-precision documents, we measure the Levenshtein distance at the character level to identify subtle misreadings (e.g., '8' vs 'B').

3. Audit Trail Reporting

Generates a "Committee-Ready" report in reports/ following every benchmark, providing a timestamped record of model performance — essential for tracking model health under production load.

⚖️ License

Distributed under the Apache License 2.0. See LICENSE for more information. Architecture inspired by the llm_eval framework for professional ML auditing standards.