Class-Incremental SAR Ship Detection and Classification via Context-Robust Exemplar Replay and Multigranularity Knowledge Distillation

Abstract

Class-incremental learning based on convolutional neural networks for synthetic aperture radar (SAR) targets has garnered wide attention recently. Unlike incremental SAR target classification, incremental SAR target detection and classification requires the model to detect and classify old targets while learning current ones, which is challenging due to the absence of old targets in the incremental large-scale data. To address the aforementioned issues, this article proposes a class-incremental SAR ship detection and classification method, which combines exemplar replay (ER) and knowledge distillation (KD) to facilitate the learning of the incremental model. Specifically, we propose a context-robust ER method to retain precise scene context between replayed old targets and clutter in the incremental data. By fully leveraging the scene characteristics of the ship targets, a context-robust ER method combines a local context-aware strategy with sea–land segmentation to pinpoint regions with the highest likelihood of being sea areas. Then, it replays exemplars within these regions, effectively reducing context-bias issues. In addition, a multigranularity KD method is introduced, which transfers the knowledge learned by the old detector to the current detector progressively. The multigranularity KD method combines an object mask strategy with spatial-channel attention mechanisms to constrain the current detector to focus on the most important information from the old detector. Experiments conducted on the SRSDD-v1.0 dataset indicate that our method achieves satisfactory performance in incremental detection and classification of SAR ships.