Cross Domain Few-Shot Learning for Fine-Grained SAR Ship Recognition

Domain adaption (DA) is one of the key problems of full domain perception using various sensors. As an effective perception technique, synthetic aperture radar (SAR) object recognition is attracting more and more attention. SAR ship recognition often needs a lot of labeled training samples. Therefore, transferring knowledge from other domain is helpful and necessary for comprehensive perception systems. DA aims to directly modify the feature space of the source and target domains to reduce the discrepancy of their distributions. However, it often needs enough samples for training and assumes that the target domain classes are consistent with the source domain, making it unable to categorize unseen classes. Few-shot learning (FSL) is a promising approach, allowing the categorizing of new, previously unknown classes in the target domain. It is used to generate predictions on the target class data by transferred knowledge (i.e., meta-knowledge from optical images) obtained from the source class data, but the domain shift has not been fully considering. Moreover, ships contain various types with subtle differences in SAR images. As a result, it must cope with new unseen fine-grained classes as well as domain shift simultaneously. To address the above composite challenges, a cross domain few-shot learning for fine-grained SAR ship recognition method called CDFS-FG is proposed, which enhancing the capabilities of SAR ship recognition. In addition, unlike conventional domain adaptation methods, Attention-Transferred Structural Pyramid Network (ATSP) is presented to construct a multi-scale attention representation for improving transferability and deriving rich contextual information by exploiting fine-grained features. Then, an enhanced semantic-aware representation is further constructed in FSL procedure, fully exploring the label correlation across tasks as well as semantic relatedness of fine-grained objects. Extensive experiments demonstrate it achieves state-of-the-art performance, which is more suitable for SAR ships recognition tasks.