Evolving classifiers with background suppression transformer for open-set long-tailed class-incremental remote...

Class-incremental learning (CIL) has recently received increasing attention in the remote sensing scene classification field, owing to its ability to quickly learn new class knowledge while retaining old class knowledge. Nevertheless, real-world remote sensing scene data is often accompanied by a long-tail distribution phenomenon and the occurrence of unknown classes, which seriously restricts the superior performance of the existing CIL models. This paper explores a more prevalent and challenging open-set long-tailed class-incremental remote sensing scene classification (OSLT-CIRSSC) task. The existing CIL methods face the following problems when generalizing them to the above OSLT-CIRSSC task: (1) Ineffectiveness of trained feature backbone for remote sensing data with complex background interference and long-tail distribution; (2) Significant confusion between known and unknown classes caused by the classifiers' decision boundary bias. To address the issues mentioned above, we propose an effective evolving classifier with the aid of the background suppression transformer (EC-BST) framework for OSLT-CIRSSC. Specifically, a long-tailed Transformer with adaptive background suppression is designed to focus more on foreground salient features of remote sensing scene data, which effectively enhances the adaptability of the trained feature backbone for long-tailed distribution in remote sensing scenes. Then, a simple classifier prediction confidence and uncertainty-based open-set recognition module is proposed to effectively evaluate class sources and avoid significant confusion between known and unknown classes. Finally, the above classifiers are continuously evolved by fusion features derived from class centroids, which effectively enhances the inter-class separability and intra-class compactness of feature embeddings and further improves the generalization of open-set recognition. Extensive experiments on three representative remote sensing datasets demonstrate the superiority of the proposed EC-BST framework in comparison to state-of-the-art long-tail CIL and open-set CIL methods.