Class-Incremental Recognition of Objects in Remote Sensing Images With Dynamic Hybrid Exemplar Selection

Class-incremental learning (CIL) aims to preserve the knowledge of former classes during the updating process of category information. Existing CIL methods often employ a singular exemplar selection strategy to reserve representative information about former classes. However, the information reserved by a singular strategy is incomprehensible, making it difficult to incrementally recognize various types of remote sensing objects from continuous data streams. To solve this problem, a dynamic hybrid exemplar selection (DHES) strategy is proposed for class-incremental recognition of remote sensing objects. In DHES, the selected exemplars incorporate both center and boundary information of former classes to mitigate the erasure of previous knowledge. During the incremental learning process, the sample proportion of hybrid exemplars is dynamically adjusted based on feature distribution to strengthen the stability of the model. Moreover, a heterogeneous prototype-based learning framework is proposed to establish a direct correlation between exemplar selection and classification prediction, thus further enhancing the informativeness of the selected exemplars. Extensive experiments on the MSTAR and FGSCR-42 datasets demonstrate that the proposed method can achieve state-of-the-art performance.