Datasets:

racineai
/

ocr-pdf-degraded

dataset_info:
  features:
    - name: page_filename
      dtype: string
    - name: pdf_filename
      dtype: string
    - name: image
      dtype: image
      struct:
        - name: bytes
          dtype: binary
        - name: path
          dtype: string
    - name: ocr_text
      dtype: string
    - name: params
      dtype: string
license: apache-2.0
task_categories:
  - text-generation
language:
  - fr
  - en
tags:
  - military
  - defense
size_categories:
  - 10K<n<100K

OCR-PDF-Degraded Dataset

Overview

This dataset contains synthetically degraded document images paired with their ground truth OCR text. It addresses a critical gap in OCR model training by providing realistic document degradations that simulate real-world conditions encountered in production environments.

Purpose

Most OCR models are trained on relatively clean, perfectly scanned documents. However, in real-world applications, especially in the military/defense sector, documents may be poorly scanned, photographed in suboptimal lighting conditions, or degraded due to environmental factors. This dataset aims to:

Enable the training of more robust OCR models that can handle imperfect document inputs
Establish a standardized benchmark for evaluating OCR performance under various degradation conditions
Bridge the gap between lab performance and real-world deployment for document processing systems

Domain Focus

This first iteration focuses specifically on military/defense sector documents. These documents:

Contain specialized terminology and formatting
Often include tables, diagrams, and structured information
May include mission-critical information where accurate OCR is essential
Represent a sector where document digitization processes may not always be ideal

Dataset Creation Process

The dataset was created through a systematic process of degrading clean PDF documents:

The process includes:

Starting with clean military/defense PDF documents
Extracting individual pages
Performing OCR on the clean pages to establish ground truth text
Applying various degradation effects to simulate real-world conditions
Recording both the degraded images and the corresponding degradation parameters

Degradation Parameters

The dataset includes various degradation types:

Noise: Random pixel noise at different intensities
Lighting: Uneven illumination effects with varying intensities and positions
Perspective: Distortions simulating non-flat document captures
Artifacts: Lines, spots, and other common scanner/camera artifacts
Image Quality: Variations in blur, brightness, contrast, and JPEG compression

Each image in the dataset includes specific parameter values, allowing for targeted evaluation and training.

Usage Examples

This dataset is ideal for:

# Example: Loading and using the dataset
from datasets import load_dataset
import json

dataset = load_dataset("racineai/ocr-pdf-degraded", split="train")

# Access a sample
sample = dataset[0]

# Get the degraded image
image = sample["image"]

# Get the ground truth OCR text
text = sample["ocr_text"]

# Access degradation parameters (for targeted training/evaluation)
params = json.loads(sample["params"])
noise_level = params["noise_level"]

print(noise_level)

Limitations and Future Work

Current iteration focuses only on military/defense documents
Further domain expansion planned for legal, medical, and financial sectors
Future versions may include handwritten text degradations
Working on expansion to include multi-page document context

Citation

If you use this dataset in your research, please cite:

@misc{racineai_ocr_pdf_degraded,
  author = {RacineAI},
  title = {OCR-PDF-Degraded: Synthetically Degraded Documents for Robust OCR},
  year = {2025},
  url = {https://huggingface.co/datasets/racineai/ocr-pdf-degraded}
}

License

Apache 2.0