Sample‐Core Class Incremental Learning for Radar Target Recognition

ABSTRACT Radar is a crucial target detection device in both defence security and civilian applications. To address the common requirement of continuous learning, radar systems require the capability to progressively acquire new classes as data streams arrive. In this context, for the task of identifying ground and low‐altitude moving targets using narrowband radar, this paper proposes a class incremental learning method. Initially, we employ wavelet transform techniques for clutter suppression and extract signal features, such as variance and radar cross‐section, to form feature vectors. These feature vectors are short in length and occupy minimal memory footprint, making them suitable for class incremental learning methods based on rehearsal. By focusing on fully utilising samples, we introduce a dynamic temperature scaling method that simultaneously adjusts both distillation loss and classification loss to enhance the model's knowledge extraction capability. Furthermore, to address the issue of class imbalance in class incremental learning, we adopt a uniform sampling strategy and relevant methods from class imbalance learning. Experimental validation on a measured radar echo dataset demonstrates the superiority of our method.