改进YOLOv8n的海洋垃圾目标检测算法：WG-YOLO

计算机工程与科学

改进YOLOv8n的海洋垃圾目标检测算法：WG-YOLO

周冬梅，王飞成，吴小所

(兰州交通大学电子与信息工程学院，甘肃兰州 730070）

出版日期:2025-06-12 发布日期:2025-06-12

Improved YOLOv8n's marine waste target detection algorithm：WG-YOLO

ZHOU Dongmei, WANG Feicheng, WU Xiaosuo

(Lanzhou Jiaotong University, School of Electronic and Information Engineering, Lanzhou,730070, China)

Online:2025-06-12 Published:2025-06-12

摘要/Abstract

摘要： 针对海洋环境下对垃圾污染物的检测精度低，小目标漏检，遮挡以及参数量大的问题，提出了WG-YOLO算法。通过CSP策略与通道洗牌方式提出S-HG Block替换C2f模块，减少了梯度信息丢失的同时，促进了对多尺度特征信息的提取。引入改进后的ESimAM注意力机制，通过生成三维注意力权重和残差连接，不仅增强了主导特征，还保留了特征较弱的特征。在Neck部分，提出了小目标特征增强金字塔网络，将富含小目标信息的低层特征融合到高层增强多尺度表示，又通过三分支通道的CSPO(CSP-OmniKernel)网络细化小目标特征。使用轻量化共享卷积检测头来优化原检测头，在大幅降低参数量的同时，保持了模型的精度。WG-YOLO模型在TrashCan数据集上mAP50达到了68.1%，mAP50:95达到了48.7%，较YOLOv8n分别提高了5.2%和2.7%。

关键词: YOLOv8, 小目标检测, 海洋垃圾, SimAM注意力机制

Abstract: The WG-YOLO algorithm is proposed to solve the problems of low detection accuracy, small target leakage, occlusion and large number of parameters in the Marine environment. Through CSP strategy and channel shuffling, S-HG Block can reduce the loss of gradient information and promote the extraction of multi-scale feature information. The introduction of the improved ESimAM attention mechanism not only enhances the dominant features by generating three-dimensional attention weights and residual connections, but also retains the weaker features. In the Neck section, a small target feature-enhanced pyramid network is proposed to fuse low-level features rich in small target information into high-level enhanced multi-scale representation, and refine small target features through the CSPO(CSP-OmniKernel) network with three branched channels. Using the lightweight shared convolution detection head to optimize the original detection head maintains the accuracy of the model while substantially reducing the number of parameters. The WG-YOLO model achieved 68.1% of mAP 50 and mAP 50:95 reached 48.7% on the TrashCan dataset, up 5.2% and 2.7% over YOLOv8n.

Key words: YOLOv8, Small target detection, marine waste, SimAM attention mechanism

周冬梅, 王飞成, 吴小所. 改进YOLOv8n的海洋垃圾目标检测算法：WG-YOLO[J]. 计算机工程与科学.

ZHOU Dongmei, WANG Feicheng, WU Xiaosuo. Improved YOLOv8n's marine waste target detection algorithm：WG-YOLO[J]. Computer Engineering & Science.

[1]	蒲小莉, 赖惠成, 高古学. BF-YOLO：基于YOLOv8改进的小目标检测算法[J]. 计算机工程与科学, 2025, 47(8): 1425-1436.
[2]	梁佳杰, 徐慧英, 朱信忠, 王舒梦, 刘子洋, 李琛. 基于改进的YOLOv8n海洋动物目标检测算法：DPSC-YOLO[J]. 计算机工程与科学, 2025, 47(4): 695-705.
[3]	张永智, 何可人, 戈珏. 改进YOLOv7网络在低空遥感图像目标检测中的应用[J]. 计算机工程与科学, 2024, 46(7): 1269-1277.
[4]	胡昭华, 王长富, . 改进Faster R-CNN的遥感图像小目标检测算法[J]. 计算机工程与科学, 2024, 46(6): 1063-1071.
[5]	邓姗姗, 黄慧, 马燕. 基于改进Faster R-CNN的小目标检测算法[J]. 计算机工程与科学, 2023, 45(5): 869-877.
[6]	吴湘宁, 贺鹏, 邓中港, 李佳琪, 王稳, 陈苗. 一种基于注意力机制的小目标检测深度学习模型[J]. 计算机工程与科学, 2021, 43(1): 95-104.
[7]	高星1，刘剑飞1，郝禄国2，董琪琪1. 基于YOLOv3算法的训练集优化和检测方法的研究[J]. 计算机工程与科学, 2020, 42(1): 103-109.