摘要
对于许多任务而言,收集注释良好的图像数据集来训练深度学习算法成本过高且耗时,而仅在渲染图像训练的模型通常无法推广到真实图像。针对上述问题,无监督域自适应算法试图在2个域之间映射一些表示或提取域不变的特征,将2个域映射到共同的特征空间。本文结合源域的有标签数据和目标域的无标签数据,提出了基于生成对抗网络(GAN)架构的无监督域自适应方法。方法使用鉴别模型,无需权重共享、对抗损失和辅助分类任务,以无监督的方式学习从一个域到另一个域的变换。对抗鉴别的无监督域自适应方法能有效减少训练域和测试域分布之间的差异,减轻域移位的有害影响,并显著地提高识别率。实验结果证明对抗鉴别方法比其他域自适应方法更有效且更简单,扩充样本的同时提高了网络的泛化性能。
Collecting well-annotated image datasets to train deep learning algorithms is prohibitively expensive and time consuming for many tasks,and models trained purely on rendered images often fail to be generalized to real images.Regarding the issues above,the unsupervised domain adaptive algorithm attempts to map some features that represent or extract domain invariance between two domains,mapping the two domains to a common feature space.Considering the labeled data of the source domain and the unlabeled data of the target domain,an unsupervised domain adaptation method based on generative adversarial network(GAN)architecture is proposed,which uses the discriminant model,without weight sharing,adversarial loss and auxiliary classification tasks,and learns the transformation from one domain to another in an unsupervised manner.The unsupervised domain adaptive method for adversarial discrimination can effectively reduce the difference between the training and test domain distributions,mitigating the harmful effects of domain shifts,and significantly improve the recognition rate.The experimental results prove that the adversarial discriminant method is more effective and simpler than other domain adaptive methods,expanding the sample and improving the generalization performance of the network.
作者
赵文仓
袁立镇
徐长凯
Zhao Wencang;Yuan Lizhen;Xu Changkai(College of Automation and Electronic Engineering,Qingdao University of Science&Technology,Qingdao 266061)
出处
《高技术通讯》
EI
CAS
北大核心
2020年第7期698-706,共9页
Chinese High Technology Letters
基金
国家自然科学基金(61171131)
山东省重点研发计划(2013YD01033)
国家留学基金委项目(201608370049)资助。
关键词
深度学习
无监督
域自适应
生成对抗网络(GAN)
辅助分类任务
deep learning
unsupervised
domain adaptation
generative adversarial network(GAN)
auxiliary classification task
