Dynamically Weighted Prototypical Learning Method for Few-Shot SAR ATR

Automatic target recognition (ATR) holds a crucial position in synthetic aperture radar (SAR) image interpretation. Despite deep-learning advancements having significantly propelled SAR ATR, addressing the challenge of target recognition with a few training data remains a vital concern in SAR applications. Two main issues still exist: 1) in few-shot SAR ATR, the depth and width of convolutional neural network (CNN)-based models are limited, which restricts its modeling capacity, and thus extracting discriminative generalized features remains challenging; and 2) with only a few labeled SAR images, the resultant class distribution is biased due to the intraclass diversity and interclass similarity of SAR samples, which degrades the recognition performance. To address these challenges, in this letter, we propose a novel dynamically weighted prototypical learning (DWPL) method. First, to extract discriminative generalized features from SAR images, we propose a new convolutional transformer network with great capacity to capture long-range dependencies of local features, together with an effective random task augmentation strategy. Second, in consideration of intraclass diversity and interclass similarity, a dynamically weighted prototypical module (DWPM) is designed to adaptively assign weights to the few labeled samples that have varying discriminative information. This enables the model to effectively explore the hidden features in a few samples. Through experiments conducted on the moving and stationary target acquisition and recognition (MSTAR) dataset, our method achieves recognition accuracies of 97.22% and 92.01% for three-way five-shot and three-way one-shot SAR ATR tasks in standard operating conditions (SOC), revealing significant and robust recognition performance.