Open AccessAttention Generating Target Images with Labels for Unsupervised Domain Adaptation
Author(s) -
Shuai Fu,
YiFan Ye,
Jing Chen
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1982/1/012064
Subject(s) - computer science , artificial intelligence , domain (mathematical analysis) , mechanism (biology) , domain adaptation , matching (statistics) , set (abstract data type) , adaptation (eye) , generative grammar , class (philosophy) , machine learning , adversarial system , field (mathematics) , unsupervised learning , mathematics , mathematical analysis , philosophy , statistics , physics , epistemology , classifier (uml) , pure mathematics , optics , programming language
Domain adaptation has recently become a popular and successful field of solving the dependence of the deep learning model on the training set. Previous works proposed unsupervised learning methods based on Generative Adversarial Networks (GANs), which solved the mixed target domain labels problem caused by the neglect of class-level matching, by generating target pairs. And these well-designed models have successfully achieved great performance on different datasets. Moreover, due to the recent widespread use of attention mechanism, we intuitively introduce attention mechanism to the previous model. Our new structure is based on GANs containing attention mechanism and well-designed upampling module, which make our model more robust. Demonstrated by extensive experiments, our new model outperforms the previous works on several standard domain adaption datasets.