A Deep Evidential Recognition Method with Geometry-Aware Features for Radar Target Rejection and Classification

In real-world scenarios, radar automatic target recognition systems face the challenge of simultaneously classifying in-library targets, and rejecting out-of-library (OOL) targets. Typically, most existing methods rely on vanilla convolution neural networks (CNNs) to extract features and adopt softmax-based or prototype-based classifiers for decision-making. However, fixed sampling grids make vanilla CNNs inflexible in adapting to the geometric structure variations in OOL targets, which leads to biased features. Besides, traditional classifiers tend to produce overconfident predictions for OOL samples, hampering effective rejection. To solve these problems, we propose a deep evidential recognition method with geometry-aware features, termed GeoDER. It consists of an adaptive geometric convolutional network (AGCNet) and a redirected evidential deep learning (REDL) framework. Firstly, AGCNet aims to model the local structure more accurately by adaptively adjusting convolution kernel shape, thereby dynamically capturing geometric anomalies in OOL targets. Then, the REDL framework leverages the geometric-aware features extracted by AGCNet to jointly infer predictive probabilities for classification and uncertainty for rejection. Specifically, the redirection loss is proposed to enhance the rejection capability, addressing the tendency of conventional evidential deep learning to prioritize classification over rejection during optimization. Experimental results on the measured dataset demonstrate that the proposed GeoDER achieves superior rejection and classification performance. Specifically, it attains 86.01% AUROC and 83.25% OSCR, with relative gains of 15.35% and 13.84% over the maximum softmax probability baseline, respectively.