Semi-supervised domain adaptation via convolutional neural network

Abstract

Semi-supervised visual domain adaptation is devoted to adapting a model learned in source domain to target domain where there are only a few labeled samples. In this paper, we propose a semi-supervised cross-domain image recognition method which unifies the feature learning and recognition model training into a convolutional neural network framework. Based on a few labeled samples and massive unlabeled samples in the source and target domains, we specially design three branches for class label, domain label and similarity label prediction which simultaneously optimizes the network to generate image features that are domain invariance and inter-class discriminative. Experimental results demonstrate that our method is effective for learning robust cross-domain image recognition model, and achieves the state-of-the-art performance on the widely used visual domain adaptation benchmark.