SAR Incremental Learning via Generative Adversarial Network and Experience Replay
摘要
Incremental learning is a viable approach for addressing the recognition problem of SAR images in data stream scenarios. However, it may suffer from catastrophic forgetting due to insufficient exposure to old category data during training. This paper proposes the integration of a generative adversarial network (GAN) into the iCaRL model to mitigate catastrophic forgetting by generating samples from previously learned categories. Experimental results on our SAR dataset demonstrate that our approach significantly enhances the recognition performance of the iCaRL model.