A Method for Sorting High-Quality Fresh Sichuan Pepper Based on a Multi-Domain Multi-Scale Feature Fusion Algorithm

Area of Science:

• Agricultural Engineering
• Computer Vision
• Food Science

Background:

• Post-harvest selection of high-quality Sichuan pepper is crucial for agricultural production.
• Existing methods struggle with varying pepper postures and maturity levels.
• Accurate visual analysis is needed for efficient sorting.

Purpose of the Study:

• To develop an automated visual system for high-quality Sichuan pepper selection.
• To improve the accuracy and efficiency of post-harvest processing.
• To enhance producer profitability through better quality control.

Main Methods:

• Proposed multi-scale frequency domain feature fusion module (MSF3M) and multi-scale dual-domain feature fusion module (MS-DFFM).
• Developed a MultiDomain YOLOv8 network for segmentation and classification of Sichuan pepper.
• Implemented a selection method based on average local pixel value difference.

Main Results:

• MultiDomain YOLOv8-seg achieved 88.8% mAP50 for fresh Sichuan pepper segmentation (5.84 MB model size).
• MultiDomain YOLOv8-cls reached 98.34% accuracy in Sichuan pepper maturity classification.
• The MultiDomain YOLOv8 model demonstrated higher accuracy and a lighter structure than the baseline YOLOv8.

Conclusions:

• The MultiDomain YOLOv8 model significantly enhances post-harvest processing efficiency for Sichuan pepper.
• The developed system effectively reduces misjudgments in quality selection.
• This technology offers substantial benefits for agricultural applications and producer profits.