A Radar HRRP Rejection Method Based on Deep Ordinal Rejector with Perturbation Generation

In real-world applications, radar automatic target recognition systems face a major challenge in the presence of out-of-library (OOL) targets, which are not collected in the databases. Existing methods have investigated on pseudo-OOL. sample generation to train a well-performed OOL target rejector. However, it is impossible to generate all OOL samples due to the diversity and unpredictability of OOL targets. Therefore, we propose a novel unsupervised method based on sample generation, named Deep Ordinal Rejector with Perturbation Generation (DORPG), which eases the requirement of synthesized samples and ranks the samples based on their similarity to training samples. It consists of a support region-aware perturbator, a convolutional encoder and an ordinal rejector, which are optimized in an end-to-end manner. Specifically, the designed perturbator generates samples that demonstrate slight sample-level difference and adjacent feature positions to form negative sample pairs with original training samples. Then the encoder is responsible for feature extraction of these negative sample pairs. Considering the multi-modality of HRRP, the developed ordinal rejector uses the minimum Mahalanobis distance as a rejection criterion and aims to enlarge the distance between negative sample pairs. Thus, DORPG achieves the more compact rejection boundary by suppressing the distribution of IL samples. Experimental results demonstrate the effectiveness of the DORPG in addressing the HRRP OOL target rejection problem.