Pairwise attention network for cross-domain image recognition

Abstract

In recent years, the domain adaption has received wide attention from machine learning communities because of differences in data distribution or the lack of training data in a practical machine learning task. In this work, we propose a Pairwise Attention Network (PAN for short) for addressing cross-domain image recognition task. In this model, different local features and the global-feature are concatenated to obtain different attention estimators, and then they are combined to get the attention map. In this way, we can focus on the important parts of an image, and ignore the irrelative regions. Moreover, attention consistency is also embedded in PAN to make sure consistent interest regions in the same class. Besides, to improve the feature discrimination, an embedding discriminative subspace is learned where it maps positive sample pairs aligned in a hypersphere and negative sample pairs separated. Extensive experimental results on the MNIST-USPS, office, and Visda-2017 datasets demonstrate that PAN can outperform state-of-the-art methods in terms of average accuracy.