基于复合注意力机制的显著目标检测方法  

Salient Object Detection Method Based on Composite Attention Mechanism

作　　者：胡成才 陈雨扬 杜欣霖 汪俊辉 陈奕霖 龚津南 HU Chengcai;CHEN Yuyang;DU Xinlin;WANG Junhui;CHEN Yilin;GONG Jinnan(School of Computer Science,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学计算机学院,四川成都610225

出　　处：《软件导刊》2025年第4期1-11,共11页Software Guide

基　　金：国家自然科学基金项目(62003058);四川省自然科学基金项目(2021YJ0165)。

摘　　要：随着多模态信息融合在计算机视觉任务中的普及,RGB-T(可见光和热红外图像)显著性检测作为一种重要的跨模态任务受到了广泛关注。现有的RGB-T显著目标检测模型大多采用卷积运算,构建复杂的多尺度融合结构实现跨模态信息集成,但其固定的感受野和权重共享机制难以有效捕捉多模态图像中的长距离依赖关系。为此,提出一个基于自注意力和交叉注意力机制的显著目标检测模型。该模型由3个关键模块组成:自注意力特征增强模块(SFEM)、多模态特征融合模块(MFFM)和特征细化模块(FRM)。SFEM混合多尺度卷积和多头自注意力机制,对提取的RGB特征和热成像特征进行增强;MFFM通过应用交叉注意力机制对所增强的特征进行多模态融合,同时连接上层特征;FRM以VGG为主干网络,通过聚合初级与深层特征来缓解深度神经网络退化问题,同时补充细节特征。实验结果显示,该模型在定量和定性比较方面均取得了显著效果。With the popularization of multimodal information fusion in computer vision tasks,RGB-T(visible light and thermal infrared imag‐es)salient object detection has received widespread attention as an important cross-modal task.Existing RGB-T salient object detection mod‐els mostly rely on convolution operations and complex multiscale fusion structures to integrate cross-modal information.However,their fixed receptive fields and weight sharing mechanisms make it difficult to capture long-range dependencies in multimodal images effectively.There‐fore,we propose a salient object detection model based on self-attention and cross-attention mechanisms,which consists of three key mod‐ules:Self-Attention Feature Enhancement Module(SFEM),Multimodal Feature Fusion Module(MFFM),and Feature Refinement Module(FRM).Specifically,SFEM integrates multiscale convolution and multi-head self-attention mechanisms to enhance extracted RGB and ther‐mal imaging features.MFFM applies cross-attention mechanisms to fuse the enhanced features across modalities while connecting upper-level features.FRM,based on a VGG backbone network,alleviates the degradation problem of deep neural networks by aggregating primary and deep features while supplementing detailed features.Extensive experiments demonstrate that the proposed network achieves excellent results in both quantitative and qualitative comparisons.

关 键 词：显著目标检测 多尺度特征 注意力机制 编码器—解码器 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]