Hybrid Reasoning Network With Class-Oriented Hierarchical Representation for Few-Shot SAR Target Recognition

The rise of deep learning has furnished a potent boost for the rapid development of automatic target recognition (ATR) in synthetic aperture radar (SAR) imagery. The existing SAR ATR methods can achieve impressive results with the great many labeled samples available. However, in real SAR application scenarios, the acquisition of quite a few SAR samples is costly or sometimes infeasible. Thus, SAR target recognition under the condition of sample scarcity is a fundamental issue to be overcome urgently. In this article, we put forward an ATR method named hybrid reasoning network with class-oriented hierarchical representation (HRNCHR) to achieve SAR target recognition with limited training samples. First, we develop a feature extraction model with local–global information concurrent refinement mechanism (LGICRM), which aims to simultaneously refine diverse features at both local and global levels from limited samples. Then, a hierarchical representation space that can learn hierarchical relationships between categories by utilizing diverse features learned from the feature extraction model is established, which is convenient to generalize to few-shot SAR target recognition tasks with the aid of category hierarchical relationship. Finally, a hybrid reasoning strategy is presented, which affords to fuse the reasoning results of instance level and prototype level to promote the accuracy of a decision-making system. Extensive evaluation experiments on the publicly released moving and stationary target acquisition and recognition (MSTAR) dataset and OpenSARship dataset illustrate that the proposed method surpasses many state-of-the-art SAR ATR methods.

I. Introduction

As an active microwave detection system, a synthetic aperture radar (SAR) plays an indispensable role in various application fields on the account of its unique operating characteristics of all-weather, all-day, and high resolution. Automatic target recognition (ATR) is a basic application in the SAR community. Yet, it appears to be one of the most challenging tasks for SAR image interpretation [1]. After decades of unremitting exploration and efforts, scholars have made proud breakthroughs in SAR target recognition [2], [3], [4]. Broadly speaking, these significant achievements can be grouped into three main streams, i.e., template-based [5], model-based [6], and learning-based one [7], [8], [9]. Among them, the template-based method infers the class attribute of the target through calculating the similarity between the template of each category and the sample to be recognized. The model-based method is dedicated to constructing a sophisticated physical model for each category of the target, which can effectively characterize the properties of the target. Different from the first two categories of methods, a learning-based method no longer extracts the features of the target in the image domain, but mines the discriminative information of the target in the predefined transformation domain, whereas, no matter in which domain, the existing methods rely severely upon handcrafted features, thereby often encountering challenges in complex SAR application scenarios.