软件导刊2025，Vol.24Issue(6)：65-71,7.DOI:10.11907/rjdk.241409
基于风格嵌入与协同学习的多目标域自适应方法
Multi-Target Domain Adaptation Approach Based on Style Embedding and Co-Learning
丁美荣 1刘庆龙 1郎济聪 1卓金鑫1
作者信息
- 1. 华南师范大学 软件学院,广东 佛山 528225
- 折叠
摘要
Abstract
Multi objective domain adaptation aims to explore how to extract knowledge from a single source domain and apply it to multiple tar-get domains,including how to ensure that the extracted domain invariant features have better category discriminability,and how to consider the correlation between a single source domain and multiple target domains.Therefore,a multi-objective domain adaptive model(SECL)based on style embedding and collaborative learning is proposed.Firstly,by changing the feature statistics of the samples,data diversity can be enhanced;Secondly,through collaborative learning of dual classifiers,optimized pseudo labels are generated for model self training,and large interval classification is used to adjust the distribution between different categories,thereby improving the model's generalization ability;Finally,the training strategy of course learning is adopted to guide the model to adapt to simple domains first,and then gradually transition to more complex domains,making the target domain adaptation process smoother.The experiment shows that the average classification accuracy of this method on Office-Home,Office31 dataset,and ImageCLEF-DA dataset is improved by 2.9%,1.1%,and 1%compared to D-CGCT.关键词
域自适应/多目标/图像分类/伪标签/自训练Key words
domain adaptation/multi-target/image classification/pseudo-labels/self-training分类
信息技术与安全科学引用本文复制引用
丁美荣,刘庆龙,郎济聪,卓金鑫..基于风格嵌入与协同学习的多目标域自适应方法[J].软件导刊,2025,24(6):65-71,7.
