ITD-YOLO: An Improved YOLO Model for Impurities in Premium Green Tea Detection

Area of Science:

• Agricultural Engineering
• Computer Vision
• Food Science

Background:

• Manual sorting of impurities in tea is labor-intensive, inefficient, and costly.
• Existing automated sorting models lack the hardware performance and suitability for industrial deployment.
• Improving the quality of premium green tea necessitates efficient and accurate impurity removal.

Purpose of the Study:

• To develop a lightweight algorithm for detecting and sorting impurities in premium green tea.
• To enhance sorting efficiency and reduce labor intensity in tea processing.
• To provide a technological foundation for sophisticated impurity classification in tea.

Main Methods:

• A custom dataset with four impurity categories was created for training and evaluation.
• Various YOLOv8 models were evaluated, with YOLOv8n selected as the base model.
• Focaler_mpdiou loss function was chosen and applied to YOLOv8m as a teacher model.
• Model pruning and knowledge distillation were employed to create a lightweight yet accurate model.

Main Results:

• The optimized lightweight model demonstrated improvements in precision (P), recall (R), and mean Average Precision (mAP).
• A significant increase in frames per second (FPS) was achieved, indicating higher processing speed.
• The model achieved reduced computational complexity (GFLOPs) and fewer parameters, making it suitable for deployment.
• The model showed generalization ability on black tea samples.

Conclusions:

• The proposed lightweight algorithm offers a viable solution for automated impurity detection and sorting in tea.
• The developed model enhances sorting efficiency and reduces manual labor in the tea industry.
• This research contributes to the advancement of automated quality control in tea processing.