Open Set Radar HRRP Recognition Using Confidence through Neural Weight Proximity

High-resolution range profiles (HRRP) have become increasingly important in radar automatic target recognition (RATR) due to their ability to capture detailed structural features of targets. In practical RATR scenarios, the operational environment is open and dynamic, often exposing models to targets belonging to previously unseen categories. This underscores the necessity for open-set recognition (OSR) methods that can effectively distinguish both known and unknown targets. Theoretically, there exists a distributional discrepancy between the feature representations of known and unknown categories in deep learning models. By designing a suitable confidence calculation method that leverages this discrepancy, it becomes possible to accurately identify whether a given sample belongs to a known or an unknown category. Inspired by the Neural Collapse theory, we observe that deep features representing known categories tend to cluster tightly around the weight vectors of the final layer, whereas features from unknown categories are typically more scattered and lie farther away. Building on this observation, we propose a confidence measure that is determined by the proximity of a feature to the weight vectors of the final classification layer. Samples with confidence values below a defined threshold are classified as unknown, which forms the fundamental mechanism of our open-set detector. Our approach is versatile, making it applicable to a variety of network architectures and adaptable to different feature distance metrics. Extensive experiments conducted on HRRP data confirm the efficacy of our proposed method for open-set recognition, demonstrating its ability to reliably distinguish between known and unknown targets under diverse conditions.