交叉注意互学习的跨领域重识别实验设计  

Design on Experiment of Cross Attention Mutual Learning for Unsupervised Cross-domain Re-identification

作　　者：云霄 褚菲 张晓光 汪群群 YUN Xiao;CHU Fei;ZHANG Xiaoguang;WANG Qunqun(School of Information and Control Engineering,China University of Mining and Technology,Xuzhou 221116,China)

机构地区：[1]中国矿业大学信息与控制工程学院,徐州221116

出　　处：《实验科学与技术》2023年第3期19-28,共10页Experiment Science and Technology

基　　金：中国矿业大学教改项目(2021YB20);江苏省高教学会评估委员会课题(2021-C158);国家自然科学基金(61902404);教育部产学合作协同育人项目(202002109018,202002177005,202002017012);教育部高等学校自动化类专业教学指导委教改项目(202123);中国矿业大学“动力中国•课程思政”示范项目(2021KCSZ31Y)。

摘　　要：针对视频智能监控中跨领域行人重识别的研究热点,将交叉注意互学习的无监督跨领域行人重识别设计为研究性实验现场硬件教学项目,观察重要区域特征关注度和网络伪标签互监督两方面对模型跨领域能力的影响。设计多尺度交叉组合注意力机制,分析其在通道和空间上对重要区域特征关注度的增强,以及对区分度不高特征关注度的减少。在此基础上设计多尺度交叉组合注意力互学习实验网络,采用硬标签互学习的方式更新训练过程,消除错误伪标签对模型性能的影响。教学实践表明,该实验项目有助于提升学生动手能力和独立思考问题的能力,为科研成果转化为教学实验提供了借鉴。Aiming at the research hotspot of cross-domain person re-identification(Person Re-ID)in video intelligent surveillance,the cross attention mutual learning for unsupervised cross-domain person re-identification is designed as a research experimental teaching project,and the influence of both the attention of important regional features and the mutual supervision of network pseudo tags on the cross-domain ability of the model is observed.A multi-scale cross combination attention mechanism is designed,and its enhancement of attention to important regional features in channel and space,as well as the reduction of attention to features with low discrimination is analyzed.On this basis,a multi-scale cross combination attention mutual learning experimental network is designed,and the training process is updated by hard label mutual learning to eliminate the impact of false labels on the performance of the model.Teaching practice shows that the experimental project helps to improve students’practical ability and independent thinking ability,and provides a reference for the transformation of scientific research achievements into teaching experiments.

关 键 词：行人重识别 无监督学习 跨领域自适应 注意力机制 互学习 

分 类 号：TP957.52[自动化与计算机技术]