Dual-modal edible oil impurity dataset for weak feature detection

Area of Science:

• Food Science
• Computer Vision
• Machine Learning

Background:

• Edible oil production can introduce solid impurities, posing food safety risks.
• Manual detection of these impurities is labor-intensive and less accurate.
• Existing datasets lack comprehensive coverage for edible oil impurity detection.

Purpose of the Study:

• To address the absence of suitable datasets for edible oil impurity detection.
• To introduce a novel dual-modal dataset for enhanced impurity identification.
• To facilitate the development of advanced computer vision models for food safety.

Main Methods:

• Development of the Dual-Modal Edible Oil Impurity (DMEOI) dataset.
• Inclusion of 14,520 event and full-picture images covering five common edible oils.
• Annotation of images for four typical solid impurities, enabling single-modal and dual-modal detection.
• Application and comparative analysis of four object detection algorithms on the dataset.

Main Results:

• The DMEOI dataset provides a valuable resource for training and evaluating computer vision models.
• Demonstrated the dataset's utility through the performance comparison of four object detection algorithms.
• Established a benchmark for future research in automated edible oil quality control.

Conclusions:

• The DMEOI dataset significantly advances the field of automated solid impurity detection in edible oils.
• The dataset supports the development of more efficient and accurate food safety inspection systems.
• Public availability of the DMEOI dataset encourages further research and innovation in food quality assurance.