Spatial Distribution Learning With Multivariate Extreme Value Boundary for Radar HRRP Open set Recognition

Abstract

In real scenarios, radar systems must identify both known and unknown classes, referred to as open set recognition (OSR). However, most prototype-based OSR methods neglect the potential impact of the directional relationship between feature embeddings and prototypes on OSR performance, leading to open space risk. Besides, the superposition of diverse constraints results in the irregular shapes of known class clusters, making accurate fitting of these cluster edges a challenging task. The spatial distribution learning (SDL) with multivariate extreme value boundary (SDL-MEVB) is proposed to address the above issues. First, we design the SDL by the fully mining and utilization of the latent representation information across the entire feature space. The proposed constraint refines the orientational distribution of feature embeddings and mitigates open space risk to the greatest extent. Second, extreme value theory is utilized to model the marginal behavior of each single variable affecting the class cluster boundary. Furthermore, we analyze the mutual dependence between these variables and establish a multivariate model to accurately fit the class cluster boundary. The effectiveness of SDL-MEVB has been verified across different backbone networks using the measured HRRPs. The experimental results demonstrate that SDL-MEVB achieves State-of-the-Art OSR performance among the implemented methods.