MeatScan: An image dataset for machine learning-based classification of fresh and spoiled cow meat

Area of Science:

• Computer Vision
• Food Science
• Machine Learning

Background:

• Deep learning advances in computer vision are not yet widely applied to practical food safety inspection, particularly in resource-limited regions.
• Accurate classification of meat freshness is crucial for public health and preventing foodborne illnesses.

Purpose of the Study:

• To introduce MeatScan, a novel, high-resolution RGB image dataset for binary classification of cow meat (fresh vs. spoiled).
• To support the development and evaluation of deep learning models for automated food quality monitoring in diverse, real-world environments.

Main Methods:

• Curated 11,000 high-resolution RGB images of cow meat from Ghanaian markets and storage facilities.
• Images were labeled as fresh or spoiled based on visual cues (color, texture, surface condition) by trained collectors under natural light.
• Dataset designed for supervised learning, accommodating convolutional neural networks, transfer learning, and data augmentation.

Main Results:

• MeatScan contains 5627 fresh and 5373 spoiled meat images, reflecting real-world conditions.
• The dataset's diverse image capture environments (markets, shops, storage) introduce variability in lighting and backgrounds.
• It serves as a benchmark for assessing model robustness against variations in meat texture and environmental factors.

Conclusions:

• MeatScan addresses a critical need for structured visual data in food safety inspection, especially in low-resource settings.
• The dataset facilitates research into robust deep learning models for meat quality assessment.
• It enables evaluation of model performance under challenging, real-world conditions, bridging the gap between AI and food safety practices.