A Review on the Few-Shot SAR Target Recognition

Synthetic aperture radar (SAR) has the advantage of providing imaging capabilities throughout the day and under all-weather conditions, which makes it particularly important for Earth observation applications. Recently, the utilization of deep learning for SAR image recognition has become a crucial discipline in radar image interpretation since the deepened networks can generate the high-dimensional features and make the function fit accurately when with a large amount of training samples. However, for SAR images, the accurate annotation demands significant effort, expert knowledge, and is prone to errors due to the effect of noise. The lack of SAR-labeled data limits the application of deep neural networks, which usually need a large number of training samples. Consequently, the task of recognizing SAR targets in the scenario with a few training samples has emerged as a significant research interest and, accordingly, the few-shot target recognition technique was introduced and has shown great potential. This article provides a summary of recent advancements in few-shot SAR image target recognition. First, this article outlines the concept of few-shot learning and discusses the dataset specific to the SAR recognition field. Subsequently, it delves into a detailed categorization of methods for recognizing few-shot SAR targets, which include approaches based on the transfer learning, data augmentation, metalearning, and model-based strategies. Finally, it examines both qualitative and quantitative aspects of SAR automatic target recognition technology utilizing few-shot learning, highlights certain challenges and crucial issues that require great attention, and offers a perspective on future research opportunities.