多重注意力改进YOLOv8的密集茶芽目标识别算法OA

To tackle the challenges of low recognition accuracy for tea buds and the difficulties in achieving automation due to high bud density and significant occlusion,an enhanced Tea-YOLOv8-Pruning model aimed at optimizing the tea bud detection performance in densely populated scenes was introduced.To augment the global feature perception capabili-ties of the Tea-YOLOv8-Pruning model,the A2Net(double attention)module was incorporated into the original YOLOv8 framework.By integrating the Multiple-SEAM module into the detection head,a relationship among features across varying density regions was established,the key characteristics of unobstructed tea buds were leveraged to predict feature information of tea buds which were highly overlapped and severely occluded buds,thereby enhancing the model's detection efficacy in dense environments.Recognizing that redundant detection boxes may arise from overlapping tea buds in such scenarios,the Repulsion loss function was introduced into YOLOv8 employing an attract-and-repel strategy,which ensured that predicted boxes converged towards true boxes while adjacent non-target prediction boxes repelled each other to mitigate the rate of missed detection and false detection.The optimal model was further optimized by pruning to improve the detection speed.Performance comparison experiments conducted on a self-constructed dataset revealed that the improved Tea-YOLOv8-Pruning network achieved average metrics of 90.6%mean Average Precision(mAP),87.9%precision(P),and 87.9%recall(R),significantly surpassing those of Faster R-CNN,YOLOv8n,and YOLOv10n models.Through comparative analysis of tea bud detection in various density scenarios,indicating that en-hanced Tea-YOLOv8-Pruning network demonstrated marked advantages,particularly in high-density and complex envi-ronments where it accurately identified more tea buds with greater confidence,and the detection performance and robust-ness were stronger.