abdoelsayed/CORU

ReceiptSense: Beyond Traditional OCR - A Dataset for Receipt Understanding

🔥 News

• [2024] ReceiptSense dataset is now publicly available!
• [2024] Paper accepted and published

📖 Abstract

Multilingual OCR and information extraction from receipts remains challenging, particularly for complex scripts like Arabic. We introduce ReceiptSense, a comprehensive dataset designed for Arabic-English receipt understanding comprising:

• 20,000 annotated receipts from diverse retail settings
• 30,000 OCR-annotated images
• 10,000 item-level annotations
• 1,265 receipt images with 40 question-answer pairs each for Receipt QA

The dataset captures merchant names, item descriptions, prices, receipt numbers, and dates to support object detection, OCR, information extraction, and question-answering tasks. We establish baseline performance using traditional methods (Tesseract OCR) and advanced neural networks, demonstrating the dataset's effectiveness for processing complex, noisy real-world receipt layouts.

🎯 Key Features

✨ Multilingual Support

• Arabic-English bilingual receipts
• Real-world mixed-language content
• Complex script handling for Arabic text

📊 Comprehensive Annotations

• Object Detection: Bounding boxes for key receipt elements
• OCR: Character and word-level text recognition
• Information Extraction: Structured data extraction
• Receipt QA: Question-answering capabilities

🏪 Diverse Retail Environments

• Supermarkets and grocery stores
• Restaurants and cafes
• Clothing and retail shops
• Various geographical regions

🔧 Real-world Challenges

• Noisy and degraded image quality
• Complex receipt layouts
• Mixed fonts and orientations
• Authentic retail scenarios

📈 Dataset Statistics

Component	Training	Validation	Test	Total
Key Information Detection	12,600	3,700	3,700	20,000
OCR Dataset	21,000	4,500	4,500	30,000
Item Information Extraction	7,000	1,500	1,500	10,000
Receipt QA	-	-	1,265	1,265

Language Distribution

• Arabic: 53.6%
• English: 26.2%
• Mixed Language: 20.3%

Receipt QA Coverage

• Merchant/Payment/Date Metadata: 30%
• Item-level Information: 50%
• Tax/Total/Payment Details: 20%

🖼️ Sample Images

Sample 1	Sample 2	Sample 3	Sample 4	Sample 5

Examples of annotated receipt images showcasing the variety of formats, languages, and complex text layouts

🎯 Supported Tasks

1. 🎯 Key Information Detection

Extract essential receipt information including:

• Merchant names
• Transaction dates
• Receipt numbers
• Item lists and descriptions
• Total amounts

2. 🔍 OCR (Optical Character Recognition)

Box-level text annotations for:

• Multilingual text recognition
• Complex layout understanding
• Noisy image processing

3. 📝 Information Extraction

Detailed item-level analysis:

• Item names and descriptions
• Prices and quantities
• Categories and classifications
• Brands and packaging information

4. ❓ Receipt Question Answering

Comprehensive QA capabilities covering:

• Receipt metadata queries
• Item-specific questions
• Transaction summary questions
• Payment and tax information

📥 Download Links

🎯 Key Information Detection

• Training Set: Download (12.6K images)
• Validation Set: Download (3.7K images)
• Test Set: Download (3.7K images)

🔍 OCR Dataset

• Training Set: Download (21K images)
• Validation Set: Download (4.5K images)
• Test Set: Download (4.5K images)

📝 Item Information Extraction

• Training Set: Download (7K items)
• Validation Set: Download (1.5K items)
• Test Set: Download (1.5K items)

❓ Receipt Question Answering

• Test Set: Download (1,265 receipts with 50.6K QA pairs)

⚠️ Note: All receipt datasets have been updated to include PII-redacted versions for privacy protection.

🏆 Baseline Results

Object Detection Performance

Model	Backbone	Precision	Recall	mAP50	mAP50-95
YOLOv7	-	76.0%	85.6%	79.2%	43.7%
YOLOv8	-	74.6%	81.0%	76.1%	45.3%
YOLOv9	-	75.7%	83.4%	77.9%	46.7%
DINO	Swin-T	-	-	-	32.2% (Avg IoU)

OCR Performance

Model	CER ↓	WER ↓
Tesseract	15.56%	30.78%
Attention-Gated CNN-BiGRU	14.85%	27.22%
Our OCR Model	7.83%	27.24%
Azura OCR	6.39%	25.97%

Receipt QA Performance

Model	Precision	Recall	Exact Match	Contains
GPT-4o	37.7%	36.4%	35.0%	29.1%
Llama3.2 (11B)	32.6%	31.3%	31.6%	25.9%
Phi3.5	28.4%	29.1%	28.8%	23.7%
Internvl2 (8B)	24.2%	23.8%	23.1%	19.4%

🚀 Getting Started

Quick Start

# Install required packages
pip install datasets transformers torch

# Load the dataset
from datasets import load_dataset

# Load Receipt QA dataset
qa_dataset = load_dataset("abdoelsayed/CORU", "qa")

# Load OCR dataset  
ocr_dataset = load_dataset("abdoelsayed/CORU", "ocr")

# Load Information Extraction dataset
ie_dataset = load_dataset("abdoelsayed/CORU", "ie")

Dataset Structure

ReceiptSense/
├── Receipt/          # Key Information Detection
│   ├── images/       # Receipt images
│   └── annotations/  # YOLO/COCO format annotations
├── OCR/             # OCR Dataset
│   ├── images/      # Text line images
│   └── labels/      # Character annotations
├── IE/              # Information Extraction
│   └── data.csv     # Structured item data
└── QA/              # Receipt Question Anshwering
    ├── images/      # Receipt images
    └── qa_pairs.json # Question-answer pairs

🔬 Applications

• 💳 Expense Management: Automated expense tracking and categorization
• 📦 Inventory Management: Real-time inventory updates from receipt data
• 🏪 Retail Analytics: Customer behavior and purchasing pattern analysis
• 🤖 Document AI: Multilingual document understanding systems
• 📱 Mobile Apps: Receipt scanning and digitization applications

🤝 Comparison with Existing Datasets

Dataset	Images	Categories	Languages	Item IE	Receipt QA	Year
SROIE	1,000	4	English	✓	✗	2019
CORD	1,000	8	English	✓	✗	2019
MC-OCR	2,436	4	EN + Vietnamese	✓	✗	2021
UIT	2,147	4	EN + Vietnamese	✓	✗	2022
ReceiptSense	20,000	5	Arabic + English	✓	✓	2024

🏛️ Ethics and Privacy

• All receipts collected with explicit user consent through the DISCO application
• Comprehensive 4-step PII redaction process implemented
• Privacy protocols strictly followed during data collection
• Independent verification and cross-checking procedures

👥 Authors

Abdelrahman Abdallah¹, Mahmoud Abdalla², Mahmoud SalahEldin Kasem², Mohamed Mahmoud², Ibrahim Abdelhalim³, Mohamed Elkasaby⁴, Yasser Elbendary⁴, Adam Jatowt¹

¹University of Innsbruck, Innsbruck, Tyrol, Austria
²Chungbuk National University, Cheongju, Republic of Korea
³University of Louisville, Louisville, USA
⁴DISCO, Cairo, Egypt

📚 Citation

If you find ReceiptSense useful for your research, please consider citing our paper:

@article{abdallah2024receiptsense,
    title={ReceiptSense: Beyond Traditional OCR - A Dataset for Receipt Understanding},
    author={Abdelrahman Abdallah and Mahmoud Abdalla and Mahmoud SalahEldin Kasem and Mohamed Mahmoud and Ibrahim Abdelhalim and Mohamed Elkasaby and Yasser Elbendary and Adam Jatowt},
    year={2024},
    journal={ACM Conference Proceedings},
    note={Comprehensive multilingual receipt understanding dataset}
}

📄 License

This dataset is released under the MIT License. See LICENSE file for details.

🔗 Links

• 📄 Paper: arXiv:2406.04493
• 🤗 HuggingFace: abdoelsayed/CORU
• 💼 DISCO App: https://discoapp.ai/
• 📧 Contact: [email protected]

---

🌟 Star this repository if you find it helpful! 🌟