Feature-Level Aspect Augmentation Network for Sparse-View Radar HRRP Recognition

In high-resolution range profile (HRRP)-based ground target recognition, collecting HRRP data with comprehensive aspect-angle coverage is often impractical. Given the strong aspect sensitivity of HRRP, sparse aspect coverage in the training set leads to an increased semantic discrepancy between seen and unseen aspects. This degrades the generalization ability of recognition methods. To improve sparse-view HRRP recognition performance, this article proposes a feature-level aspect augmentation network (FLAAN), consisting of two stages, namely feature generation and feature augmentation. By augmenting the training set with generated features that provide representations of additional aspects, the proposed method enhances the semantic association between seen and unseen aspects. In the feature generation stage, a generative model with a scattering-center-guided constraint is proposed to reduce semantic distortion in the generated features and improve their effectiveness. In the feature augmentation stage, a multiview-aware mechanism exploits the generated features to promote intraclass consistency over large aspect intervals, thereby further mitigating the impact of aspect sensitivity on recognition performance. Finally, we adopt an end-to-end joint optimization strategy to tightly couple the generative and recognition processes. In this way, the generative process is directly aligned with the discriminative objectives of the recognition model, ensuring that the generated features better support downstream recognition performance. Experimental results demonstrate that under five different sparse-aspect sampling cases, the proposed method outperforms various approaches based on few-shot learning and data augmentation, achieving a maximum of 10.19% improvement under the 72° aspect interval sampling setting.