Few-Shot SAR Target Classification Combining Both Spatial and Frequency Information

The automatic recognition of synthetic aperture radar (SAR) targets has been extensively studied in recent years. Specifically, due to the high cost of SAR image acquisition and the low occurrence probability of high-value targets, it is of great importance to identify SAR targets with only a few available images, which is referred to as few-shot SAR target classification. However, most existing solutions straightly adopt meta-learning and transfer learning methods that are originally designed for optical images, which do not consider the unique frequency information of SAR images. In this paper, we propose a novel hybrid classification network that combines both spatial and frequency information for few-shot SAR target classification. Specifically, we first train the proposed network in a source dataset, which contains a large number of related SAR images without the targets of interest. Then, the pre-trained network is fine-tuned on the target dataset consisting of only a few SAR images of interest. Compared with the conventional method based on convolutional neural networks and model-agnostic meta-learning, the proposed method can achieve superior top-1 accuracy in various settings.