A Novel Angular-Based Unsupervised Domain Adaptation Framework for Image Classification

Domain adaptation (DA) deals with the common problem of distribution mismatch between the training and the target data. A model tested on data that comes from a different source than the training data will not perform well. DA methods aim to utilize the available training and testing data to model a target domain classifier. Domain invariant features are extracted and used to minimize the distribution divergence between the source and target domains. At the same time, there exist some imperative limitations. The angle between the subspaces of the source and target domains is not reduced during distribution alignment, and the formulation of an objective function, that reduces geometric and statistical deviations, must be used. None of the DA strategies in use today address these constraints simultaneously. Therefore, we propose a novel DA framework, called angular-based unsupervised domain adaptation framework for image classification (AUDAF) to address these limitations. AUDAF first obtains the pseudo labels for the unlabeled target domain data using a simple k-Nearest Neighbor classifier (trained on source domain data) and then formulates a robust objective function to reduce domain discrepancy. The objective function can be solved efficiently in a closed form by considering a couple of projection vector matrices (one for source, the other for target). Extensive testing on benchmark DA datasets shows that AUDAF performs better than existing DA methods in terms of classification accuracy.