Attribute-Informed and Similarity-Enhanced Zero-Shot Radar Target Recognition

Although high-resolution range profile (HRRP) target recognition methods have achieved satisfactory performance in closed set recognition tasks, they are not able to recognize unseen classes that are not trained. In the real-world scenario, it is usually difficult or even impossible to collect all potential classes, especially for noncooperative targets. Aiming at this problem, we propose an attribute-informed and similarity-enhanced (AttSim) zero-shot HRRP target recognition method. We argue for using scattering center vectors as side information to learn useful knowledge from seen classes and apply it to the recognition of unseen classes, which successfully alleviates the electromagnetic semantic ambiguity problem. Additionally, an attribute-informed scattering center vector generation method is proposed to extract scattering center vectors of seen classes and to generate those of unseen classes. The scattering center vectors of unseen classes are estimated by the weighted summation of those of seen classes. The weights are chosen from the similarities of the target compositional structures between the seen and unseen classes. Exploiting the similarities, a similarity-inspired projection function is proposed to map HRRP features to scattering center embeddings. The proposed projection function is directly defined rather than learned only using seen class samples, thus is capable of mitigating the projection domain shift to a certain extent. The classification of unseen classes is then achieved by finding the most similar scattering center vectors to their scattering center embeddings. In addition, a similarity-enhanced scattering center loss function is devised to enhance the discriminability of scattering center embeddings. The proposed loss function aims to pull the scattering center embeddings close to scattering center vectors of their corresponding classes while simultaneously pushing them away from scattering center vectors of other classes, with the extent of the separation determined by the similarity between different classes. Extensive experiments verify the effectiveness of the AttSim method for zero-shot HRRP target recognition.