Learn to Recognize Unknown SAR Targets From Reflection Similarity

Abstract

In real-world recognition tasks, it is difficult to collect training samples for all types of synthetic aperture radar (SAR) targets. The existing zero-shot learning (ZSL) methods use visual features and label semantics simultaneously to ensure a high recognition rate of unknown targets. However, meaningless information could influence the stability of the reference space. The negative impact will further drop the recognition rate of the ZSL methods. To overcome this problem, a new ZSL method is proposed by only adopting reflection similarity to recognize unknown targets. The proposed method adopts a triple-part network, in which the constructor-net and the supervisor-net jointly construct the reference space. It makes the space have a good representation and distinction. The third part of the triple-part network, as an interpreter-net, provides the uniformity of the target interpretation. The experimental results on moving and stationary target acquisition and recognition (MSTAR) data set show that the reference space with good quality can be obtained by discarding meaningless information. Especially, when the samples of known targets have slightly changed, the reference space still keeps good stability. Compared with the traditional ZSL methods, the recognition reliability of the proposed method on zero-shot targets is improved.