双注意力随机选择全局上下文细粒度识别网络  

Dual-attention random selection global context fine-grained recognition network

作　　者：徐胜军[1,2] 荆扬 段中兴 李明海[1,2] 李海涛 刘福友[4] XU Shengjun;JING Yang;DUAN Zhongxing;LI Minghai;LI Haitao;LIU Fuyou(College of Information and Control Engineering,Xi'an University of Architecture and Technology,Xi'an 710055,China;Xi'an Key Labratory of Building Manufactaring Intelligent&Automation Technology,Xi'an 710055,China;Traffic Engineering Construction Bureau of Jiangsu Province,Nanjing 210024,China;CCCC Tunnel Engineering Company Limited,Beijing 100024,China)

机构地区：[1]西安建筑科技大学信息与控制工程学院,陕西西安710055 [2]西安市建筑制造智能化技术重点实验室,陕西西安710055 [3]江苏省交通工程建设局,江苏南京210004 [4]中交隧道工程局有限公司,北京100024

出　　处：《液晶与显示》2024年第4期506-521,共16页Chinese Journal of Liquid Crystals and Displays

基　　金：国家自然科学基金(No.52278125);陕西省自然科学基础研究项目(No.2023-JC-YB-532,No.2022JQ681);陕西省重点研发计划(No.2021SF-429);陕西省教育厅专项科研计划(No.20JK0721)。

摘　　要：针对细粒度图像识别任务中易忽视微小潜在性特征且外观差异细微等问题,提出一种基于双注意力随机选择全局上下文细粒度识别网络。首先,使用ConvNeXt作为主干网络,提出双注意力随机选择模块,对不同阶段提取到的特征进行通道随机选择和空间随机选择,使网络能够关注到其他潜在微小判别性特征;其次,利用全局上下文注意力模块将深层特征的语义信息融合到中间层,增强中间层定位微小特征的能力;最后,提出一种多分支损失,对中间层、深层和拼接层特征引入分类损失,结合不同分支提取到的特征,诱导网络获得多样性的判别特征。所提网络在Stanford-cars、CUB-200-2011、FGVC-Aircraft 3个公开细粒度数据集和真实场景下车型数据集VMRURS上分别达到了95.2%、92.1%、94.0%和97.0%的识别准确率,其性能相比其他对比方法有较大幅度提升。To address the difficulties of capturing the potential distinguishable features and subtle appearance differences in fine-grained image recognition tasks,dual-attention random selection global context fine-grained recognition network is proposed.Firstly,the ConvNeXt is taken as the backbone network,a dual-attention random selection module is proposed to perform channel random selection and spatial random selection on the features extracted at different stages,so that the network could focus on other potential subtle distinguishable features.Then,by using the global context attention module,the semantic information of top-level is applied to the middle-level to enhance the ability of the middle-level to locate potential subtle distinguishable features.Finally,the multi-branch loss is proposed,and classification loss is imposed on middle-level,top-level and concat-level characteristics.Combining the features extracted from different branches,the network is induced to obtain diverse distinguishable features.The network achieves the accuracies of 95.2%,92.1%,94.0%and 97.0%respectively on the three open datasets,Stanford-cars,CUB-200-2011,FGVC-Aircraft and dataset VMRURS in real scenes.The presented method in this paper greatly upgrades the comparison performance.

关 键 词：细粒度识别 ConvNeXt 双注意力随机选择 全局上下文注意力 多分支损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]