基于组卷积和膨胀卷积的轻量注意力模块  

Lightweight attention module based on group convolution and dilated convolution

作　　者：张恩淘 郝晓丽[1] 牛保宁[1] ZHANG En-tao;HAO Xiao-li;NIU Bao-ning(College of Computer Science and Technology,Taiyuan University of Technology,Jinzhong 030600,China)

机构地区：[1]太原理工大学计算机科学与技术学院,山西晋中030600

出　　处：《计算机工程与设计》2025年第2期493-499,共7页Computer Engineering and Design

基　　金：国家自然科学基金面上基金项目(62072326)。

摘　　要：为解决目前的注意力模块中参数量大、通道压缩导致信息丢失、空间信息学习不充分的缺点,提出一种基于组卷积、通道清洗和膨胀卷积的轻量注意力模块。采取组卷积和通道清洗的方式学习通道权重,能够在不压缩通道的前提下减少大量参数,使不同组之间产生交互,充分学习通道信息。采取连续的膨胀卷积,合理设置膨胀率充分且均衡的学习空间信息。通过CIFAR100和VOC 2007+2012数据集对所提模块在图像分类和目标检测中进行实验,验证其能够在较少的花费下带来较大的提升。To address the shortcomings of large parameter count,channel compression leading to information loss,and insuf-ficient spatial information learning in current attention modules,a lightweight attention module based on group convolution,channel shuffling,and dilated convolution was proposed.Using group convolution and channel shuffling to learn channel weights could reduce a large number of parameters without compressing channels,and enabling interaction between different groups to fully learn channel information.Adopting continuous dilation convolutions and setting the dilation rates reasonably could fully and balanced learn spatial information.Experiments on the proposed module in image classification and object detection were carried out on CIFAR100 and VOC2007+2012 datasets.It is verified that the proposed attention module can bring greater improvement with less cost.

关 键 词：深度学习 卷积神经网络 注意力机制 组卷积 膨胀卷积 图像分类 目标检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]