Decision Boundary Optimization for Few-shot Class-Incremental Learning

Few-shot class-incremental learning (FSCIL) is gaining prominence in real-world machine learning applications, including image classification and face recognition. Existing methods often employ parameter freezing for the backbone and classify based on metric learning. However, these methods suffer from two significant problems. Firstly, training the backbone solely on base classes limits its performance on novel classes due to information loss. Secondly, conventional metric-based strategies for prototype generation tend to introduce confusion in decision boundaries during few-shot tasks. To address these challenges, we propose a novel approach called Decision Boundary Optimization Network (DBONet) for few-shot class-incremental learning. To tackle the first issue, DBONet incorporates an augmentation feature extractor along with a corresponding loss function. This augmentation feature extractor combines samples from different categories to capture richer features. For the second issue, we leverage limited sample representativeness information by introducing the Prototype Generation Module (PGM) into DBONet, enabling the generation of more representative prototypes. The prototypes produced by PGM significantly contribute to the accurate delineation of decision boundaries. Furthermore, we exploit intra-class information to enhance classification precision. Extensive experiments on CIFAR100, miniImageNet, and CUB200 datasets demonstrate that our proposed approach achieves new state-of-the-art results.