基于注意力融合的语义分割网络  

Semantic Segmentation Using Attention Fusion

作　　者：李旭升 何良华[1] 程舸帆 LI Xu-sheng;HE Liang-hua;CHENG Ge-fan(College of Electronic and Information Engineering of Tongji University,Shanghai 201800,China)

机构地区：[1]同济大学电子与信息工程学院计算机科学与技术系,上海201800

出　　处：《电脑知识与技术》2021年第1期1-3,21,共4页Computer Knowledge and Technology

基　　金：国家自然科学基金(项目号:61772369)的支持;国家自然科学基金联合项目(项目号:U1809206)的支持;上海市科委国际合作项目(项目号:17411953100,19490712800)的支持;上海市技术带头人项目支持(项目号:19XD1434000)。

摘　　要：为赋予语义分割网络在给定空间位置下选择性强调整体信息或细节信息的能力,提出了一种注意力融合算法,本算法在空洞空间金字塔池化(ASPP)的基础上融合胶囊网络中动态路由算法。首先,以骨干网络输出作为输入,经过多条并行空洞卷积支路得到不同尺度的特征图。然后,在每一条空洞卷积支路的后面增添一条评估支路来评测该条空洞卷积支路单独分割的能力。最后,对各个评估支路的输出进行注意力路由算法从而对各空洞卷积分配权重。在PascalVOC 2012和Cityscapes两个数据集上,提出的模型在各组实验中均能提升1个百分点以上,并通过可视化注意力图表明,提出模型能够根据上下文信息对各空洞卷积支路进行有侧重的反向传播。To enable semantic segmentation networks to learn to emphasize the overall picture or the focal details at given spatial lo⁃cation,Attention Fusion is proposed,which combines Atrous Spatial Pyramid Pooling and dynamic routing of Capsule Network.Firstly,parallel atrous branches with different ratios are applied on feature map extracted by backbone network,Secondly,an eval branch for each atrous branch is used to output a score map in order to assess the ability of each atrous branch to segment on its own.Finally,attention routing assigns weights for atrous branches by computing attention maps between these score maps.Experi⁃ments conducted on Pascal VOC 2012 and Cityscapes,and the performances are boosted by over 1 percent in all settings.Along with visualization of attention maps,it is shown that the proposed model is capable of emphasizing atrous branches according to giv⁃en context.

关 键 词：语义分割 卷积神经网络 空洞空间金字塔池化 胶囊网络 动态路由 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]