融合混合注意力的轻量化柑橘成熟度检测算法OA

To address the challenges faced in citrus detection on fruit trees in natural environments,such as difficulty in distinguishing fruit maturity,severe occlusion from branches and leaves,high model complexity,and resource deployment limitations,this paper proposes a lightweight citrus maturity detection algorithm for fruit trees,YOLO-HiP,based on an improved YOLOv11.First,an improved HGNetv2-L network is used as the backbone,combined with a hierarchical feature extraction strategy,significantly enhancing the model's capability to analyze complex scenes while effectively reducing computational complexity and resource consumption.Next,a lightweight hybrid attention module,C2PSA_iRMB,is designed.By integrating the C2PSA mechanism with the iRMB module,the computational cost is optimized,and the ability to process long-range information is enhanced,improving the module's flexibility and computational efficiency.Finally,a C3k2_PConv module is constructed,further improving spatial feature extraction efficiency by reducing redundant computations and memory access.Experimental results show that YOLO-HiP achieves 94.3%mAP50,an improvement of 4.7%over the original model,with only 5.1M parameters(a 45.7%reduction),a computational load of 13.9 GFLOPs(a 34.7%reduction),and a frame rate of 227.4 FPS(a 25.1%increase).This model significantly reduces model size while ensuring detection accuracy,providing an innovative and feasible solution for platforms with limited computational resources,such as citrus-picking robots and other embedded systems.