Few-Shot Class-incremental SAR Target Recognition via Orthogonal Distributed Features

As synthetic aperture radar (SAR) imaging technology continues to evolve, the growing repository of SAR images depicting diverse types of observed targets has sparked rising interest in SAR target incremental recognition techniques. However, most existing SAR target incremental recognition algorithms typically require an ample amount of training data. In urgent scenarios such as emergency response and disaster relief, there may be a necessity to identify targets for which a substantial amount of data has not been previously accumulated. Algorithms designed for general scenarios often fail to achieve satisfactory performance in such situations. To tackle the aforementioned issues, this paper presents a few-shot incremental recognition algorithm for SAR targets based on orthogonal distributed features. Specifically, an orthogonal distribution optimization method for features is designed, which not only mitigates the feature confusion in few-shot incremental learning, but also reserves space for features of potential unseen classes. A random augmentation method for high-dimensional features is proposed to improve the overfitting problem while assisting in strengthening the boundaries between features of different classes. Furthermore, a joint decision criterion based on Euclidean distance and cosine distance is introduced, enabling the classifier to possess sufficient generalization ability and robustness in handling dynamic data. Experimental results on the MSTAR dataset show that the algorithm outlined in this paper outperforms existing methods in SAR target few-shot incremental recognition tasks, demonstrating its superior effectiveness.