Continuous Teacher-Student Learning for Class-Incremental SAR Target Identification

In this paper, we propose a class-incremental SAR target identification approach based on continuous teacher-student learning. The main challenge of class-incremental SAR target identification is catastrophic forgetting: as the learned model tend to adapt to the most recently seen new identification task, they forget what they have learned before and therefore lose performance on the tasks that were learned previously. Our method aims at introducing the teacher model, which can utilize data from tasks so far, to prevent the student model from catastrophic forgetting. For each task, the teacher model learn to capture the knowledge contained in the tasks by now. When a new task is presented, the student model is encouraged to learn from the teacher model so that the information on which the previous task relied is retained. At the same time, we also make the student model to review its own knowledge to further alleviate catastrophic forgetting. The evaluation of continuous SAR target recognition task shows that this method reduces forgetting effect.