面向特征生成的无监督域适应算法  

Unsupervised Feature-Level Domain Adaptation with Generative Adversarial Networks

作　　者：吴子锐 杨之蒙 蒲晓蓉[1] 徐杰 曹晟[1] 任亚洲 WU Zirui;YANG Zhimeng;PU Xiaorong;XU Jie;CAO Sheng;REN Yazhou(School of Computer Science and Engineering,University of Electronic Science and Technology of China Chengdu 611731;Institute of Electronic and Information Engineering,University of Electronic Science and Technology of China Dongguan Guangdong 523808)

机构地区：[1]电子科技大学计算机科学与工程学院,成都611731 [2]电子科技大学广东电子信息工程研究院,广东东莞523808

出　　处：《电子科技大学学报》2022年第4期580-585,607,共7页Journal of University of Electronic Science and Technology of China

基　　金：国家自然科学基金(61806043);四川省科技计划(2021YFS0172,2020YFS0119);广东省基础与应用基础研究基金(2020A1515011002);中央高校基本科研业务费(ZYGX2021YGLH022)。

摘　　要：针对无标签高维图像分类问题,常用的深度网络在无标签的情况下难以产生好的分类结果。为此,提出一种面向特征生成的无监督域适应模型(Feature-GAN),它以一种无监督的方式在特征层面学习从一个域到另一个域转换,将源域图像特征映射为目标域图像特征并保持标签信息,生成的带标签特征可用于目标域特征的分类训练。该模型在复杂图像域适应上避免了图像本身的生成过程,而专注于特征生成,易训练且稳定性高。实验表明,该方法可以广泛应用于复杂图像分类的场景,相比于传统基于样本生成的无监督域适应算法,该算法在精确度、收敛速度以及稳定性上均有提高。For the classification problem of unlabeled high-dimensional images, the commonly used deep neutral networks have difficulty in producing good classification results in the unlabeled datasets. This paper proposes an unsupervised feature-level domain adaptation with generative adversarial networks(Feature-GAN),which learns the feature level transformation from one domain to another in unsupervised manner. It maps the source domain image features to the target domain image features and keeps the label information, and these generated labeled features can be used to train a classifier adapted to the target domain features. This model avoids the generation process of the image itself in the complex image domain adaptation problem and focuses on feature generation. The model is easy to train and has high stability. Experiments show that the proposed method can be widely applied to complex image classification scenarios, and it outperforms traditional sample generation-based unsupervised domain adaptation algorithms in terms of accuracy, convergence speed, and stability.

关 键 词：生成对抗网络 图像分类 迁移学习 无监督域适应 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]