YOLOESD-基于改进YOLOv8的海洋垃圾图像检测  

YOLOESD-Marine Litter Image Detection Based on Improved YOLOv8

作　　者：李翠 王姣[1] LI Cui;WANG Jiao(School of Software,Dalian Jiaotong University,Dalian,Liaoning 116021,China)

机构地区：[1]大连交通大学软件学院,辽宁大连116021

出　　处：《计算技术与自动化》2025年第1期41-45,58,共6页Computing Technology and Automation

摘　　要：海洋垃圾不仅严重威胁海洋动物的健康及其栖息地,由其释放的有毒物质进入食物链后对人类身体健康同样造成消极影响。由于海洋图像受到光照投影的影响,并且垃圾的尺寸通常较小,以往的目标检测算法对海洋垃圾的检测性能并不理想,因此提出了一种基于YOLOv8网络模型的改进算法(YOLOESD),该算法共有三个改进点:首先,采用Stemblock模块替换了模型的初始卷积,在减少模型参数量的同时,提高模型检测的精确度;其次,融合高效多尺度注意力模块(EMA),有效减少了模型的漏检和误检问题;最后,在原模型的头部额外增加一个小目标检测头,提高模型对小尺度目标的敏感度。实验结果表明,改进后的YOLOv8网络模型与原网络模型相比,漏检情况得到明显改善,mAP@0.5达到90.8%,精度提高了3.6个百分点;YOLOESD网络模型的检测效果优于原网络模型及经典的网络模型。Marine litter is a serious threat to the health of marine animals and their habitats,and it also has a negative impact on human health through the release of toxic substances that enter the food chain.The performance of previous target detection algorithms for marine litter is not satisfactory due to the fact that marine images are affected by light projection and the size of the litter is usually small.Therefore,this paper proposes an improved algorithm(YOLOESD)based on the YOLOv8 network model,which has three improvement points.Firstly,the initial convolution of the model is replaced by the Stemblock module.While reducing the number of model parameters,improve the detection accuracy of the model.Secondly,we have integrated the efficient multi-scale attention module(EMA).Effectively reducing the problem of missed and false alarms in the model.Finally,an additional small target detection head is added to the head of the original model to improve the model's sensitivity to small-scale targets.The experiments results show that the improved YOLOv8 network model has significantly improved leakage detection compared to the original network model,the mAP@0.5 reached 90.8%,and the accuracy is improved by 3.6 percentage points;the YOLOESD network model outperforms both the original network model and the classical network model in terms of detection.

关 键 词：目标检测 小目标检测 海洋垃圾检测 EMA注意力机制 Stem模块 YOLOESD 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]