A Class Incremental Learning Method Based on Class Rebalancing and Old Knowledge Transfer

Real-world scenarios are inherently dynamic where new objects constantly emerge. Class incremental learning (CIL), which learns to classify new classes by updating old classification model, has been proposed for object classification problems in real world. Current CIL approaches typically retain a few exemplars from old classes in a size-constrained memory and replay them when updating the model with new class data. However, such an update mode faces two serious issues: catastrophic forgetting and class imbalance, leading to the loss of discriminative knowledge among old classes and the tendency to misclassify old class data towards new classes, respectively. As a result, the performance of old classes after updating will be destroyed. In this paper, we propose a CIL method based on class rebalancing and old knowledge transfer that can alleviate catastrophic forgetting and class imbalance in CIL. In detail, the proposed method is mainly composed of two modules: feature-relation knowledge distillation (FRKD) and prototypical metric rebalancing (PMR). FRKD copes with catastrophic forgetting, transferring the learned discriminative knowledge among old classes by encouraging the updated model to reproduce similar feature topologies with the old model. PMR addresses class imbalance by employing prototypical metric learning to adjust the model tendency towards new classes, thereby enhancing the distinguishability between old and new class features. With the FRKD and PMR techniques, our method adequately retains the performance on old classes when learning new classes. Experiments on benchmark datasets show the superiority of our method compared to the state-of-the-art methods.