Hierarchical Distribution-Based Exemplar Replay for Incremental SAR Automatic Target Recognition

Abstract

Over the years, deep learning-based automatic target recognition (ATR) in synthetic aperture radar (SAR) imagery has made remarkable progress on the assumption that the target category library is immutable. However, the target category library will continue to expand over time in real-world scenarios, so the ATR model should be updated to acquire reasoning capabilities for subsequent acquired targets. If the current ATR model is updated only with samples from subsequent acquired targets, it will perform poorly on old categories of targets, which is known as catastrophic forgetting. Due to timeliness and storage resource constraints, it is not feasible to retraining the ATR model from scratch with all training samples. Therefore, how to select representative exemplars that can characterize the entire distribution is a crucial issue for incremental SAR target recognition. In this article, we present an ATR method named hierarchical distribution-based exemplar replay (HDER) to achieve incremental SAR target recognition. Specifically, we first propose a hierarchical class distribution-based exemplar selection strategy, which can afford to select those diverse and representative exemplars based on the maximum covariance and minimal similarity criteria. To recall more the knowledge of old categories from replay exemplars, we then present a multilevel distillation mechanism that takes into account replaying the rich knowledge contained in the feature level and classification level. Finally, a multitask loss is formulated to train the whole recognition network, to ensure that the recognition network is competent to continuously generalize on subsequent acquired targets while maintaining the reasoning ability for old categories as much as possible. Extensive experiments on moving and stationary target acquisition and recognition dataset reveal that the proposed HDER surpasses some State-of-the-Art incremental recognition methods under different scenarios.