Exploring a Discriminative Metric Network for Few-Shot SAR Target Classification

Few-shot learning aims to learn a classifier for image classification when only a few training samples are available. Recent researches have shown that using local descriptors instead of holistic image representations can provide more informative features. However, local descriptors only capture local information within an image while disregarding global contextual information, leading to decreased accuracy in synthetic aperture radar (SAR) target classification. To address this issue, we propose a novel discriminative metric network (DMnet) for few-shot SAR target classification. The DMnet comprises global-level metric relationship and local-level metric relationship. Among them, the local-level metric relation is computed by an image-to-class metric module that measures the similarity of local descriptors. Simultaneously, we introduce global pooling to obtain the global information for each local descriptor, reducing the influence of local background noise. Subsequently, a prototype-based metric module is utilized to calculate the similarity of descriptors. To better integrate these two types of metric relations, we design an adaptive metric fusion module to achieve more discriminative metric performance. Experimental results on the MSTAR dataset demonstrate the effectiveness of the proposed DMnet.