Lightweight improvement algorithm for target detection of Pu'er tea harvesting robotic arm based on YOLOv8

Abstract

To tackle the challenges of recognition difficulties and constrained computational resources in Pu'er tea intelligent harvesting, this research develops an optimised, resource-efficient object detection algorithm built upon the YOLOv8n architecture for detecting tender shoots of Pu'er tea. The methodology incorporates three primary enhancements: first, replacing the standard Conv module with the Adown down-sampling component enhances detection precision, significantly boosts processing speed, and minimises model complexity; second, modifying the detection head to the LADH configuration cuts down parameter volume, further streamlining the model; third, integrating the AFGC attention mechanism refines detection accuracy. Experimental outcomes reveal that the optimised model achieves a 0.7% increase in mean average precision (mAP), accelerates detection speed by 482.9 FPS, and reduces model size by 1.9 MB compared to the baseline YOLOv8n. This work provides a technical foundation for advancing intelligent harvesting systems tailored for Pu'er tea cultivation.