Infrared Small-Sample Target Recognition With Content-focused Domain Adaption Network

The domain adaptation learning paradigm provides an approach for the small-sample infrared image recognition task by introducing a source domain with sufficient visible optical samples. It aims to learn domain-shared features from the source samples and transfer them to the infrared image domain. However, the performance will suffer from the domain shift problem which is in part caused by the inductive preference of the learning model on image style instead of its content. To address the above issues, this paper develops a novel content-focused domain adaption network. We propose two novel inductive biases from both intra-domain and cross-domain perspectives to make the learning model more focused on the content. First, an infrared intra-domain content-focused sub-network based on retinex theory is designed so that the style accounting for the image noise will not influence the intra-domain induction. Next, we construct a cross-domain content-focused sub-network to alleviate the attention on the image style. Thanks to these two networks, the overall domain adaptation model will pay more attention to the robust discriminative feature of image content rather than domain styles. Extensive experiments demonstrate the effectiveness of our algorithm. In particular, the recognition accuracy of our method can exceed the baseline method by up to 8%.