Frequency Characteristics Guided Network for Few-Shot SAR Target Recognition

Contemporary research in synthetic aperture radar (SAR) automatic target recognition (ATR) reveals that few-shot learning algorithms can attain exceptional classification accuracy through training paradigms employing several hundred to thousands of sample inputs. However, existing methods ignore the frequency characteristics in radar images and only rely on the similarity of pixel descriptors for target recognition. To overcome this limitation, this paper presents frequency characteristics guided network (FCGN), an architecture explicitly developed for SAR ATR scenarios with limited training samples. First, we propose a frequency-separated feature extractor (FSFE), which enriches the frequency characteristics of the target. In addition, FCGN further incorporates a frequency-domain sample expander (FDSE), a dedicated component for generating spectrally congruent pseudo-samples that enhance support set heterogeneity, ultimately refining class separation boundaries in the latent representation space. Finally, we propose an adaptive frequency-domain matcher (AFDM). AFDM calculates the inter-sample frequency-domain consistency through selected frequency components, and the network synthesizes the pixel consistency and frequency-domain consistency to discriminate the samples. Rigorous evaluation on the moving and stationary target acquisition and recognition (MSTAR) dataset demonstrate that the proposed method surpasses current approaches.