基于改进YOLOv5s的漂浮垃圾目标检测算法  

Floating waste detection algorithm based on improved YOLOv5s

作　　者：张岩松 高茂庭[1] ZHANG Yan-song;GAO Mao-ting(College of Information Engineering,Shanghai Maritime University,Shanghai 201306,China)

机构地区：[1]上海海事大学信息工程学院,上海201306

出　　处：《计算机工程与设计》2025年第3期910-917,共8页Computer Engineering and Design

基　　金：国家重点研发计划基金项目(2020YFC1511901)。

摘　　要：针对漂浮垃圾检测存在小目标特征丢失引起的漏检误检和模型过于复杂导致部署困难的问题,提出一种基于改进YOLOv5s的目标检测算法YOLOv5s-PBW。采用改进的PB-FPN特征融合结构,在不影响检测精度的情况下,降低模型的复杂度;引入双层路由注意力机制(BRA),关注关键信息,提高对小目标特征的提取能力;使用Wise-IoU边框损失函数,专注于训练普通质量的锚框,提高模型的泛化能力。实验结果表明,改进后的模型相比原始YOLOv5s算法,参数量减少了62%,权重文件减少了61%,mAP@0.5提高了1.2%,适合无人水面船的漂浮垃圾实时检测。Aiming at the problems of the floating waste detection algorithm,such as missed detections and false detections caused by the loss of small target features and deployment difficulties caused by overly complex models,an improved YOLOv5s detection algorithm YOLOv5s-PBW was proposed.An improved PB-FPN feature fusion structure was used to reduce model complexity without affecting detection accuracy.The bi-level routing attention mechanism(BRA) was added to focus on key information,so as to improve the ability to extract small target features.The bounding box regression loss function Wise-IoU was used to focus on ordinary-quality anchor boxes and increase the generalization ability of the model.Experimental results show that the improved model reduces parameters by 62%,reduces weight files by 61% and increases mAP@0.5 by 1.2% compared to the original YOLOv5s algorithm,which is conducive to further real-time detection of floating waste on unmanned surface vehicle.

关 键 词：深度学习 目标检测 YOLOv5s 漂浮垃圾 特征融合 双层路由注意力机制 损失函数 

分 类 号：TP319.4[自动化与计算机技术—计算机软件与理论]