An annotated grain kernel image database for visual quality inspection.

Area of Science:

• Agricultural Science
• Computer Vision
• Data Science

Background:

• Visual quality inspection of grain kernels is crucial for trade, storage, and agriculture.
• Existing datasets lack the scale and detail required for advanced machine vision applications.
• Accurate grain assessment impacts food security and economic value.

Purpose of the Study:

• To introduce GrainSet, a large-scale, machine vision-based database for grain kernel visual quality inspection.
• To provide comprehensive annotations including morphology, physical size, weight, and damage categories.
• To establish a benchmark for deep learning models in grain analysis.

Main Methods:

• Collected over 350,000 single-kernel images of wheat, maize, sorghum, and rice from diverse global regions.
• Utilized a custom-built device with high-resolution optic sensors for surface data capture.
• Incorporated expert annotations for sampling information and quality attributes, including damage and unsound categories.
• Implemented a standard deep learning model for initial classification benchmarking.

Main Results:

• Developed GrainSet, a database exceeding 350K annotated single-kernel images.
• Captured detailed surface information and expert-verified quality assessments for four major cereal grains.
• Demonstrated the potential of deep learning models for grain classification using the dataset.

Conclusions:

• GrainSet provides a valuable resource for advancing machine vision in grain quality assessment.
• The database will support research in automated inspection, storage guidance, and smart agriculture.
• Facilitates improved accuracy and efficiency in grain trade and management.