Unified Discriminant and Distribution Alignment for Visual Domain Adaptation

Abstract

In visual understandings, images taken from different cameras usually have different resolutions, illumination, poses, and background views that lead to domain shift. Besides labeling these data is an expensive operation. These problems lead to the need for unsupervised domain adaptation (UDA), in which training and testing data are not drawn from the same distribution, and labels are not available in the target domain. This paper presents an improvement for unsupervised domain adaptation in transfer learning using a unified discriminant and distribution alignment (UDDA). The existing method of UDA only utilized unsupervised PCA as the dimensionality reduction process before being added to the joint objective function consisting of distribution discrepancy minimization and regularization. However, the label in the source domain has been utilized by some works (i.e., joint geometrical and statistical (JGSA)) to use the supervised method LDA and show good improvement. Nevertheless, LDA has some drawbacks that is sensitive to noise and outlier in square operations and only take part in global information. The contribution of this paper is to add local discriminant information to the proposed UDA model by adopting locality-sensitive discriminant analysis (LSDA) that strengthens the between-class and within-class discriminant in the source domain. In this method, within-class and between-class graphs will be computed, and summed up with within-class and between-class scatter matrices before being embedded into the joint domain adaptation framework. Comparing the proposed method with the state-of-the-art techniques in objects and digital datasets, it shows our UDDA improved the average accuracy by 3.43% and 7.13% compare with the second and third highest results respectively.