Incremental SAR Target Recognition via Spatial Coverage Maximization-Based Exemplar Replay

In real application environments, synthetic aperture radar (SAR) systems continuously acquire new types of targets, which requires that the recognition model should have the ability to continuously accommodate new categories. However, existing recognition methods often lead to performance degradation on old categories with the addition of new categories, that is, catastrophic forgetting. In view of this, this letter proposes a method called spatial coverage maximization-based exemplar replay (SCMER) to achieve class incremental SAR target recognition. First, a spatial coverage maximization-based exemplar selection is proposed, which resorts to hierarchical clustering method to partition the feature space and selects exemplars from old categories through different criteria in the hierarchical space. Then, a categories relationship constraints-based knowledge recall mechanism is developed to recall knowledge from selected exemplars as much as possible to alleviate catastrophic forgetting. Finally, the recognition model is incrementally updated through multi-task learning loss. Experiments on two SAR image classification datasets show the effectiveness of our method compared to other state-of-the-art methods.