广东海洋大学学报2025，Vol.45Issue(2)：109-117,9.DOI:10.3969/j.issn.1673-9159.2025.02.014
融合多种注意力机制和Wise-IoUv3的水下目标检测算法
Underwater Target Detection Algorithm with Multiple Attention Mechanisms and Wise-IoUv3
摘要
Abstract
[Objective]To solve the problems of imaging blurring and low detection accuracy in complex background of underwater target images,this study proposes an underwater target detection algorithm which integrates multiple attention mechanisms and Wise-IoUv3.[Method]Firstly,a multi-scale feature enhancement mechanism was designed in the backbone network,omni-dimensional dynamic convolution(ODConv)was used to replace some convolutions,and an efficient-scale attention mechanism(EMA)was introduced to enhance the backbone network's ability to extract features of blurred targets and small targets.Secondly,the spatial pyramid pooling-fast(SPPF)module was improved by adding an average pooling branch to supplement spatial information and enhance global context awareness.In addition,the lightweight BiFormer attention mechanism was integrated into the two branches to reduce the model's computational complexity and enhance the detection performance of small targets.Then,in the prediction stage,the original loss function was replaced by Wise-IoUv3 to balance the model training results of different quality images.Finally,the original detection head was replaced by the dynamic detection head(DynamicHead)to enhance the scale perception,spatial perception and task perception ability of the detection head and improve the accuracy of object position recognition.[Result]The experimental results on RUOD and URPC datasets show that the improved algorithm increases the mean average precision by 3.6%and 1.7%respectively,compared with the benchmark methods(especially YOLOv8n);the model's parameter count and computational cost are reduced by 0.26×106 and 0.4 GFLOPs,respectively.[Conclusion]The experimental results show that the proposed method reduces the missed-detection and false-detection of blurred targets and small targets in complex situations,thus improving the detection and maintaining the model's lightweight design.关键词
水下目标检测/多尺度特征增强机制/多尺度注意力机制/全维动态卷积/Wise-IoUv3Key words
underwater target detection/multi-scale feature enhancement mechanism/multi-scale attention mechanism/omni-dimensional dynamic convolution/Wise-IoUv3分类
信息技术与安全科学引用本文复制引用
肖振久,高凯歌,李士博..融合多种注意力机制和Wise-IoUv3的水下目标检测算法[J].广东海洋大学学报,2025,45(2):109-117,9.基金项目
辽宁省高等学校基本科研项目(LJKMZ20220699) （LJKMZ20220699）
辽宁工程技术大学学科创新团队资助项目(LNTU20TD-23) （LNTU20TD-23）
