文献类型：期刊文献

中文题名：水下目标的多尺度上下文感知检测模型

英文题名：Multi-Scale Context-Aware Detection Model for Underwater Target

作者：王峻韬[1];郑红[1];陆元军[2];徐贤[1];吴丽娟[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237;[2]印孚瑟斯技术(中国)有限公司杭州分公司,杭州310056

年份：2026

卷号：52

期号：2

起止页码：276

中文期刊名：华东理工大学学报(自然科学版)

外文期刊名：Journal of East China University of Science and Technology

收录：;北大核心:【北大核心2023】；

基金：上海市2024年度“科技创新行动计划”(24BC3200500,24BC3200300)。

语种：中文

中文关键词：水下目标检测;深度学习;注意力机制;下采样;特征提取

外文关键词：underwater object detection;deep learning;attention mechanism;downsampling;feature extraction

摘要：针对传统模型无法有效处理水下复杂环境噪声、目标尺度变化大、且无法平衡模型大小和精度的问题,本文提出了MSCA-UODA(Multi-scale Context-Aware Underwater Object Detection Algorithm)模型,其设计的上下文增强下采样模块CEADown(Context Enhanced ADown)在降低模型参数量的同时能有效捕获上下文信息,减少了下采样过程中水下环境噪声的影响;同时,提出一种基于双路径部分连接的多尺度特征提取模块CSP-MSPF(Cross Stage Partial-Multi-Scale Partial Feature),并使用单头注意力机制(Single-Head Self-Attention,SHSA)来改进C2PSA,提高了模型的多尺度特征提取能力。实验表明,相较于基准模型,MSCA-UODA模型在数据集URPC2020和DUO的mAP50分别提升了2.0个百分点和1.1个百分点,参数量下降了12.01%,且综合性能优于目前主流的目标检测模型。
To address the limitations of traditional models in handling complex underwater environmental noise,large variations in target scale,and the trade-off between model size and accuracy,the MSCA-UODA(Multi-scale Context-Aware Underwater Object Detection Algorithm)was proposed.The model includes a context-enhanced downsampling module,CEADown(Context-Enhanced ADown),which effectively reduces model parameters,captures contextual information efficiently,and mitigates underwater environmental noise.Additionally,it introduces a multiscale feature extraction module based on dual-path partial connection,named CSP-MSPF(Cross Stage Partial-Multiscale Partial Feature),and incorporates the SHSA(Single-Head Self-Attention)mechanism to enhance the C2PSA module,thereby improving the model's multi-scale feature extraction capability.Experimental results show that on the URPC2020 and DUO datasets,MSCA-UODA improved mAP50 by 2.0 percentage points and 1.1 percentage points,respectively,compared to the baseline model,while reducing the number of parameters by 12.01%.Its overall performance surpassed that of current mainstream object detection models.