关键词: 花椒簇, 目标检测, YOLOv5, 轻量化, 深度学习

Abstract: The automation of prickly ash cluster harvesting is a crucial step towards achieving intelligent production of prickly ash and reducing production costs. Accurate and efficient detection of prickly ash clustersis an important prerequisite for automated harvesting systems. This study introduces a lightweight prickly ash cluster detection model, GGL—YOLO, based on YOLOv5s, specifically designed for this task. The model incorporates several enhancements. First, the SE attention mechanism is integrated into the shallow layers of GhostNet, while its redundant deeper layers are removed, resulting in an optimized GhostNet serving as the YOLOv5s backbone. Second, the neck of the model is improved by introducing a lighter GSConv convolution method and the C2f module. Finally, a dynamic receptive field is applied to the P3, P4, and P5 feature map outputs. Experimental results show that, compared tothe YOLOv5s model, the GGL—YOLO model achieved a 1.5% increase in the mean average precision(mAP), a 58.6% reduction in parameter volume, a 66.7% reduction in FLOPs, and a 13.2% increase in detection speed. These results confirm that the GGL—YOLO model offers a feasible solution for the fast and accurate detection of prickly ash clusters, thereby advancing the automation of their harvesting process.

Key words:  , prickly ash cluster, object detection, YOLOv5, lightweight, deep learning