Enhancing multilevel tea leaf recognition based on improved YOLOv8n.

Area of Science:

• Agricultural Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Automated tea picking requires precise tea leaf recognition for efficiency and quality.
• Deep learning shows promise in tea detection, but multilevel composite feature analysis is lacking.
• Existing methods struggle with recognizing diverse tea leaf categories and overlapping targets.

Purpose of the Study:

• To enhance the recognition accuracy of different tea leaf categories for automated picking.
• To develop a robust deep learning model capable of detecting small and overlapping tea leaf targets.
• To create an efficient and deployable solution for smart tea cultivation.

Main Methods:

• Proposed a novel method for generating overlapping-labeled tea category datasets.
• Introduced the Tea-You Only Look Once v8n (T-YOLOv8n) model for multilevel composite tea leaf detection.
• Integrated Convolutional Block Attention Module (CBAM) and Bidirectional Feature Pyramid Network (BiFPN) for feature fusion.
• Utilized CIOU and Focal Loss functions to optimize bounding box predictions.

Main Results:

• The T-YOLOv8n model achieved superior performance in detecting small and overlapping tea leaf targets.
• Achieved a precision increase from 70.5% to 74.4% and recall from 73.3% to 75.4% in mAP50 compared to YOLOv8, YOLOv5, and YOLOv9.
• Reduced computational costs by up to 19.3%, demonstrating efficiency.
• Showcased enhanced adaptability to diverse lighting and background variations.

Conclusions:

• The T-YOLOv8n model offers improved detection accuracy and computational efficiency for automated tea picking.
• The model's robustness and adaptability make it suitable for practical deployment in resource-constrained edge computing environments.
• This research contributes to smart agriculture by advancing intelligent tea leaf classification and automated harvesting, boosting tea production efficiency and sustainability.