Enhancing Multilevel Tea Leaf Recognition Based on Improved YOLOv8n

Xinchen Tang Li Tang ^* Junmin Li Xiaofei Guo

• Xihua University, Chengdu, China

In the tea industry, automated tea picking plays a vital role in improving efficiency and ensuring quality. Tea leaf recognition significantly impacts the precision and success of automated operations. In recent years, deep learning has achieved notable advancements in tea detection, yet research on multilevel composite features remains insufficient. To meet the diverse demands of automated tea picking, this study aims to enhance the recognition of different tea leaf categories. A novel method for generating overlapping-labeled tea category datasets is proposed. Additionally, the Tea-You Only Look Once v8n (T-YOLOv8n) model is introduced for multilevel composite tea leaf detection. By incorporating the Convolutional Block Attention Module (CBAM) and the Bidirectional Feature Pyramid Network (BiFPN) for multi-scale feature fusion, the improved T-YOLOv8n model demonstrates superior performance in detecting small and overlapping targets. Moreover, integrating the CIOU and Focal Loss functions further optimizes the accuracy and stability of bounding box predictions. Experimental results highlight that the proposed T-YOLOv8n surpasses YOLOv8, YOLOv5, and YOLOv9 in mAP50, achieving a notable precision increase from 70.5% to 74.4% and recall from 73.3% to 75.4%. Additionally, computational costs are reduced by up to 19.3%, confirming its robustness and suitability for complex tea garden environment. These results verify the robustness and high detection accuracy of the algorithm, making it well-suited for complex tea garden environments. These findings underscore the effectiveness of the algorithm, providing an innovative solution for intelligent tea classification and automated picking.