摘要
无监督领域自适应(Unsupervised Domain Adaptation,UDA)旨在帮助模型在跨域分布差异条件下从带标注的源域中学习到知识,以迁移至无标注的目标域。当前先进的域自适应方法主要通过直接对目标域与源域分布对齐来实现,其中特征往往被当作一个整体对象用于开展域间自适应任务,忽略了特征中的任务关联信息(域间不变、域内独特信息)与无关信息(颜色对比度、图像风格)耦合的情况,使得模型难以把握关键的特征信息,从而导致次优化。针对上述问题,提出了一种基于任务关联特征解耦网络的无监督领域自适应分类方法(Task Relevant Feature Separation Network,TRFS),通过对域间风格混合干扰下的特征与原始特征的注意力进行一致性的学习,来帮助网络提炼出与下游任务相关的特征权重,并进一步采用权重差获取任务无关特征权重,而后通过正交函数约束推远任务关联与无关特征,实现特征解耦;设计了任务特征细化解耦层,减轻配对特征与域独特特征混淆的情形,优化模型对分类判别的精度。此外,为了提升伪标签质量,引入基于记忆力银行的领域聚合伪标签生成方法,用于降低伪标签噪声。综合实验结果表明,所设计解耦模块具有良好的即插即用性,能够提升自适应方法的性能;且所提方法相比其他先进的域适应方法具有明显的优势,其中在Office-Home数据集上达到了73.6%的分类精度。
Unsupervised domain adaptation(UDA)aims to assist the model in transferring learned information from a labeled source domain to an unlabeled target domain,given cross-domain distribution discrepancies.Current advanced domain adaptation techniques rely mostly on aligning the distributions of the source and target domains.Among them,the features are frequently utilized as a global object to perform inter-domain adaptation tasks,disregarding the coupling of task-related information(inter-domain invariant,intra-domain specific information)and task-irrelevant information(color contrast,image style)in the features.The situation makes it difficult for the model to comprehend the important information of features,resulting in sub-optimization.In consideration of the aforementioned issues,we propose an unsupervised domain adaptive classification method based on a task relevant feature separation network(TRFS),which helps the network extract the downstream task-related feature weight by learning the attention consistency between the features with inter-domain style mixed interference and the original features.Further,weight subtraction is used to obtain the task-irrelevant feature weight,then the task-related and irrelevant features are further pushed away by orthogonal function constraints to achieve feature separation.The task feature refinement separation layer is designed to reduce the confusion situation of alignment features and domain-specific features,as well as optimize the model’s classification and discrimination accuracy.Comprehensive experiment results demonstrate that the designed separation module has good plug-and-play performance,which can enhance the performance of other UDA methods.And the TRFS has obvious advantages over other advanced UDA methods,which achieves a classification accuracy of 73.6%in the Office-Home benchmark.
作者
唐珺琨
张辉
张邹铨
吴天月
TANG Junkun;ZHANG Hui;ZHANG Zhouquan;WU Tianyue(School of Electrical&Information Engineering,Changsha University of Science&Technology,Changsha 410000,China;School of Robotics,Hunan University,Changsha 410000,China)
出处
《计算机科学》
CSCD
北大核心
2023年第S02期281-288,共8页
Computer Science
基金
国家重大研究计划-重点支持项目(92148204)
国家重点研发计划(2021ZD0114503)
国家自然科学基金(61971071,62027810,62133005)
湖南省杰出青年科学基金项目(2021JJ10025)
湖南省重点研发计划(2021GK4011,2022GK2011)
长沙市科技重大专项(kh2003026)
机器人学国家重点实验室联合开放基金(2021-KF-22-17)
中国高校产学研创新基金(2020HYA06006)
湖南科技创新领军人才项目(2022RC3063)。
关键词
特征解耦
任务关联
注意力机制
无监督领域自适应
图像分类
Feature decoupling
Task-relevant
Attention mechanism
Unsupervised domain adaptation
Image classification
