Domain-invariant representation learning using an unsupervised domain adversarial adaptation deep neural network

Abstract

Domain adaptation is proposed to improve the recognition performance of the domain shift or the dataset bias. The domain shift is a very common problem, which is caused by diverse factors, such as data capturing angles, illumination, and image quality existing in the natural scene image understanding. Since the domain shift leads to the feature distribution discrepancy, some solutions have been proposed to alleviate the distribution discrepancy by mapping feature spaces between source and target domains, so as to ensure the transferable features can be learned by the deep networks during the end-to-end training for the classification tasks. However, it is still a big challenge to address the domain shift when the distribution spaces are not clearly separated. Inspired by the adversarial idea, we propose a novel unified deep neural network architecture named the unsupervised domain adversarial adaptation deep neural network. It addresses the domain adaptation problem by learning domain-invariant features through mitigating the feature discriminative ability in the domain classification task alternatively by alleviating the feature distribution discrepancy in the main classification task. Therefore, in our proposed unified deep network, we integrate two main modules. One is an auxiliary task module for the domain classifier, which is trained to make sure the learned features are domain-invariant under the adversarial optimization strategy by minimizing the domain discriminative ability. The other is the module at task-specific layers to enhance the learning of the transferable features with the less distribution discrepancy by adding multiple maximum mean discrepancy constraints to map the features to reproducing kernel Hilbert spaces. The experimental results show that the features learned by our proposed unified deep neural network perform better than the features learned by previous cross-domain neural networks on classification tasks. Our proposed approach achieves the state-of-the-art performance on three cross-domain datasets: Office-31 (different capturing angles, illumination, and image quality), Office-Caltech (modified from Office-31) and a combined cross-domain digit dataset, including MNIST, USPS and SVHN (different style digits in each dataset).