文献类型：期刊文献

中文题名：结合自注意力的对抗性领域适应图像分类方法

英文题名：Adversarial domain adaptation withself-attention in image classification

作者：陈诚[1];郭卫斌[1];李庆瑜[2]

机构：[1]华东理工大学信息科学与工程学院,上海200237;[2]上海网达软件股份有限公司,上海201206

年份：2020

卷号：42

期号：2

起止页码：259

中文期刊名：计算机工程与科学

外文期刊名：Computer Engineering & Science

收录：CSTPCD;;北大核心:【北大核心2017】；CSCD:【CSCD_E2019_2020】；

基金：国家自然科学基金(61672227)。

语种：中文

中文关键词：迁移学习;领域适应;图像分类;生成对抗网络

外文关键词：transfer learning;domain adaptation;image classification;adversarial networks

摘要：作为解决数据集迁移和适应的系统性框架,领域适应在近年来发展迅速。在生成对抗网络出现以后,对抗性思想的引入为领域适应中的无监督适应问题带来了新的思路。通过研究生成对抗网络和领域适应的内在联系,类比生成对抗网络的改进方法,提出了结合自注意力模块的领域适应方法,用以弥补无法建模长距离依赖的缺陷。同时,考虑到生成对抗网络和领域适应任务上的不同,通过引入新的学习参数来改进自注意力模块,使其在分类任务上有更高的精度和健壮性。最后,在公开的领域适应数据集上的实验证实了本文方法的有效性和可行性。
Domain adaptation,as a systematic framework for addressing data set migration and adaptation,has grown rapidly in recent years.After the emergence of the generative adversarial network(GAN),the introduction of adversarial ideas has brought new ideas to the unsupervised adaptation problem in the domain adaptation.This paper compares and analyzes the intrinsic relationship between GAN and domain adaptation,and then generates an improved GAN method.A domain adaptation method combining self-attention layer is proposed to improve the defect that long-distance dependence cannot be modeled.At the same time,considering the difference in GAN and domain adaptation tasks,this paper improves the self-attention layer by introducing new learning parameters,so that it has higher precision and robustness in classification tasks.Finally,the effectiveness and feasibility of the proposed algorithm are verified by experiments on the open field adaptive dataset.