prabhats0605/scanned-images-dataset-for-ocr-and-vlm-finetuning

Dataset Card for scanned_images_dataset

This is a FiftyOne dataset containing 3,482 scanned document images across 10 diverse document categories. Designed for OCR training and Vision-Language Model (VLM) fine-tuning, this dataset features real-world scanned documents with varied layouts, scanning quality, and document types.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/scanned-images-dataset-for-ocr-and-vlm-finetuning")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

This dataset is a curated collection of scanned document images representing 10 common document types found in office, academic, and professional settings. The images exhibit real-world characteristics including varied scanning quality, different layouts, mixed content types (text, tables, images), and both printed and handwritten text.

The dataset is particularly valuable for:

• Training OCR models on diverse document layouts

• Fine-tuning Vision-Language Models for document understanding

• Developing document classification systems

• Testing robustness to real-world scanning artifacts and quality variations

• Original Source: Kaggle - Scanned Images Dataset for OCR and VLM finetuning

• Curated by: suvroo (original), Harpreet Sahota (FiftyOne port)

• Language(s): English (primary), with potential multilingual content

• License: Original license from Kaggle source

Dataset Sources

• Original Repository: Kaggle Dataset
• Dataset Download: GTS.AI Mirror

Uses

Direct Use

This dataset is suitable for:

• OCR Model Training: Train text extraction on real-world scanned documents with varied quality
• Document Classification: Develop models to classify documents by type (email, memo, form, etc.)
• Vision-Language Model Fine-tuning: Adapt VLMs for document understanding tasks
• Layout Analysis: Detect and understand document structure (headers, paragraphs, tables)
• Document Quality Assessment: Train models to handle various scanning qualities
• Multi-column Text Recognition: Practice on complex layouts (news articles, scientific papers)
• Form Understanding: Extract structured information from forms and tables
• Handwriting Recognition: Some documents contain handwritten elements
• Real-world OCR Benchmarking: Evaluate model performance on authentic scanned documents

Out-of-Scope Use

• Privacy-Sensitive Applications: Documents may contain personally identifiable information
• Production Document Processing: Without text annotations, this dataset is primarily for classification and pre-training
• Language-Specific OCR: While primarily English, multilingual content may be present but not systematically labeled
• High-Precision Applications: Scanning artifacts and quality variations may limit use in applications requiring perfect text extraction

Dataset Structure

FiftyOne Schema

Each sample in the dataset has the following structure:

{
    'id': str,                      # Unique sample ID
    'media_type': 'image',          # Always 'image'
    'filepath': str,                # Absolute path to document image
    'tags': [],                     # Empty list (no tags)
    'metadata': {                   # Image metadata
        'size_bytes': int,          # File size (typically 200KB-800KB)
        'mime_type': 'image/jpeg',  # MIME type
        'width': int,               # Image width (varies, typically 1200-2500px)
        'height': int,              # Image height (varies, typically 1500-3200px)
        'num_channels': 1,          # Grayscale (1 channel)
    },
    'ground_truth': {               # Document classification label
        'id': str,                  # Label ID
        'tags': [],                 # Empty list
        'label': str,               # Document category (see below)
        'confidence': None,         # Not applicable
        'logits': None,             # Not applicable
    }
}

Document Categories

The dataset contains 10 document categories with the following distribution:

Category	Count	Description
Email	599	Email correspondence and communications
Memo	620	Internal memorandums and office notes
Letter	567	Formal and informal letters
Form	431	Forms, applications, and structured documents
Report	265	Business and technical reports
Scientific	261	Scientific papers and research documents
ADVE (Advertisement)	230	Advertisements and promotional materials
Note	201	Handwritten and typed notes
News	188	News articles and clippings
Resume	120	Résumés and CVs
Total	3,482

Image Characteristics

• Format: JPEG
• Color Mode: Grayscale (1 channel)
• Resolution: Varies (typically 72 DPI)
• Typical Dimensions: 1200-2500 × 1500-3200 pixels
• File Size: 200KB - 800KB per image
• Orientation: Mostly portrait, some landscape

Document Features

The scanned documents exhibit:

• Multi-column layouts (news articles, scientific papers)
• Tables and forms with structured data
• Mixed content (text, images, logos)
• Varied text styles (printed, handwritten, typewritten)
• Real scanning artifacts (skew, noise, shadows, folds)
• Different scanning qualities (low to high resolution)
• Multiple fonts and sizes
• Headers, footers, and page numbers

Dataset Creation

Curation Rationale

This dataset was created to provide a diverse collection of real-world scanned documents for training and evaluating OCR systems and Vision-Language Models. The variety in document types, layouts, and scanning quality makes it valuable for developing robust document understanding systems that can handle real-world conditions.

Source Data

Data Collection and Processing

The dataset consists of scanned images from various document types commonly found in professional and academic settings. Documents were scanned at varying quality levels to represent real-world scenarios.

This FiftyOne port was created by:

1. Downloading the original dataset from Kaggle
2. Organizing images into 10 category directories
3. Using FiftyOne's ImageClassificationDirectoryTree format to import with automatic label assignment
4. Computing metadata for all images (dimensions, file size, color channels)

The directory structure maps directly to classification labels:

scanned-images-dataset-for-ocr-and-vlm-finetuning/
├── ADVE/          (230 images)
├── Email/         (599 images)
├── Form/          (431 images)
├── Letter/        (567 images)
├── Memo/          (620 images)
├── News/          (188 images)
├── Note/          (201 images)
├── Report/        (265 images)
├── Resume/        (120 images)
└── Scientific/    (261 images)

Who are the source data producers?

The documents appear to be collected from various sources representing typical office, business, and academic materials. Specific provenance information is not available from the original dataset.

Annotations

Annotation Process

The dataset includes document type classification labels derived from the directory structure. Each image is labeled with one of 10 document categories based on its location in the folder hierarchy.

No text-level annotations (OCR transcriptions, bounding boxes, or entity labels) are included. The dataset is designed for:

• Document classification tasks
• Pre-training OCR and VLM models
• Layout analysis research
• Unsupervised or self-supervised learning

Personal and Sensitive Information

⚠️ Privacy Warning: This dataset contains scanned documents that may include:

• Personal names and addresses
• Email addresses and contact information
• Company names and business information
• Academic affiliations
• Potentially sensitive correspondence

Users should:

• Review documents before use in production systems
• Implement appropriate privacy safeguards
• Consider anonymization or redaction for sensitive applications
• Comply with relevant privacy regulations (GDPR, CCPA, etc.)

Bias, Risks, and Limitations

Known Limitations

• No Text Annotations: Dataset lacks OCR transcriptions, making it unsuitable for supervised text extraction training without additional annotation
• Language Bias: Primarily English documents; limited representation of other languages
• Domain Coverage: May not represent all document types (e.g., forms from specific industries)
• Quality Variation: Scanning quality varies significantly, which could impact model training if not handled properly
• Imbalanced Categories: Category sizes range from 120 (Resume) to 620 (Memo) samples
• Privacy Concerns: Contains potentially identifiable information without systematic anonymization
• Temporal Bias: Document styles and formats may reflect specific time periods
• Geographic Bias: Document formats may be specific to certain regions

Recommendations

• Use for Classification or Pre-training: Best suited for document type classification or unsupervised pre-training
• Combine with Annotated Datasets: For OCR training, combine with datasets that include text annotations
• Privacy Review: Audit documents for sensitive information before deployment
• Augment for Balance: Apply data augmentation or resampling to address class imbalance
• Quality Filtering: Consider filtering or stratifying by scan quality for specific applications
• Validate on Target Domain: Test on your specific document types before production use
• Consider Licensing: Verify usage rights for your specific application

Citation

If you use this dataset, please cite:

Original Dataset:

@misc{scanned-images-ocr-vlm,
  author = {suvroo},
  title = {Scanned Images Dataset for OCR and VLM finetuning},
  year = {2024},
  publisher = {Kaggle},
  url = {https://www.kaggle.com/datasets/suvroo/scanned-images-dataset-for-ocr-and-vlm-finetuning}
}

FiftyOne Port:

@misc{scanned-images-ocr-vlm-fiftyone,
  author = {Harpreet Sahota},
  title = {Scanned Images Dataset for OCR and VLM finetuning (FiftyOne)},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/voxel51/scanned-images-dataset-for-ocr-and-vlm-finetuning}
}

Dataset Card Contact

For questions or issues with this FiftyOne port, please open an issue on the Hugging Face dataset repository.