Tea bud detection in complex natural environments based on YOLOv8n-RGS

摘要

Abstract To address the challenge of accurately detecting tender tea buds under natural conditions due to occlusion, uneven lighting, and missed small targets, this study proposes a lightweight detection method called YOLOv8n-RGS, based on YOLOv8n. The method focuses on small object detection in occluded environments. First, Region Attention Networks (RAN) are embedded into the backbone to adaptively enhance key region features and effectively suppress interference caused by leaf occlusion. Second, a GSConv (Group Shuffle Convolution) structure is introduced in the neck to combine the advantages of standard convolution and depthwise separable convolution, which improves multi-scale feature representation while reducing model complexity. Finally, the Slide loss function is used to dynamically adjust the weight of positive and negative samples, addressing sample imbalance in scenarios with occlusion and uneven lighting, and further improving detection accuracy. Experimental results show that, compared with the original YOLOv8n, the proposed optimized model reduces model size and computational cost by 3.2% and 4.8% respectively, and increases inference speed by 4.1%. Meanwhile, the F1 score (balanced F Score), recall, and mean average precision (mAP) are improved by 1%, 4%, and 3.1%, respectively. Compared with other mainstream lightweight models such as YOLOv4, YOLOv5n, and YOLOv7-Tiny, YOLOv8n-RGS achieves significantly better detection performance. This model provides an effective solution for high-precision bud detection and occlusion suppression in tea-picking robots.