20 March 2025 in IEEE Sensors Journal
High Resolution Range Profile (HRRP), with its concise vector representation and rich structural information, has become the primary focus of research in the field of Radar Automatic Target Recognition (RATR) in Sensors community. In actual combat situations, due to limitations in sensing conditions, acquiring data from non-cooperative targets is challenging. Consequently, identifying HRRP targets in the absence of sufficient data is a significant and urgent issue that requires immediate attention. Additionally, precise target recognition represents a pressing hurdle. To this end, this paper proposes a maksed Vision Transformer (ViT) and brownian distance covariance classifier (MTBC). The ViT is initially pre-trained through a self-supervised learning strategy of masked reconstruction, which enables the model to gain a comprehensive understanding of the intricate interconnections between neighbouring blocks in HRRP, thereby facilitating the extraction of more nuanced and discriminative features. Additionally, within the context of prototype learning in few-shot metric learning, a brownian distance covariance metric-based classifier is devised. This combines a parametric feature extractor and a non-parametric metric classifier, enabling the accurate description of the dependency between support class prototypes and query vectors. Experimental results validated on measured HRRP dataset demonstrate that the proposed method performs superiorly in few-shot HRRP target recognition task.
27 February 2025 in IEEE Transactions on Instrumentation and Measurement
In real-world scenarios, radar systems need to recognize both known and unknown classes, referred to as open set recognition (OSR). Prototype-like methods have shown significant potential in addressing this challenging task. However, prototype learning overlooks the directional relationship between feature embeddings and prototypes, leading to open space risk. In addition, strong learning for known classes can weaken the sensitivity to sample variations, impacting OSR performance of the model. To address these issues, we propose the regularization-based region learning (RBRL) method. Firstly, from the perspective of gradient propagation, RBRL demonstrates that l2-regularization can enhance the sensitivity of logit vector directions to sample variations. Unlike in closed-set scenarios, l2-regularization can effectively improve OSR performance even with sufficient training data. Secondly, RBRL minimizes the known space in both distance and angular domains by leveraging the spatial geometric relationship between prototypes and feature embeddings, mitigating open space risk. Furthermore, decision-level fusion is achieved by computing the area of the fan-shaped region in RBRL. The effectiveness of RBRL is validated on datasets comprising both measured and simulated high resolution range profiles (HRRPs). RBRL achieves state-of-the-art performance across various OSR metrics.