End-to-End Domain Adaptation Network for Cross-Domain Image Retrieval

With the intense emergence of information technology, image data becomes widely used in multiple applications, such as information retrieval. Deep supervised methods have been proven to be the most efficient approaches for image retrieval. However, most of the existing deep supervised image retrieval methods assume that the database set and query set are sampled from similar distributions, i.e., same domains, which significantly constrains their implementation in real-world applications. In this paper, we propose an end-to-end domain adaptation network to address the challenging image retrieval task, in the cases that the database set and query set come from different domains. Specifically, in our work, the images in the target domain are unlabeled, which makes the proposed method unsupervised. Moreover, we extend a traditional convolutional neural network to a novel domain adaptation network by adding a discriminator to distinguish the features generated by the feature extractor. Besides the softmax loss, three different levels of loss functions are used to improve the network’s feature extraction ability in the target domain. We conduct experiments on multiple cross-domain tasks and the results show the effectiveness and superiority of the proposed method compared to the existing approaches.