局部关系泛化表征的小样本增量学习

Few-shot learning is a booming research area in computer vision that aims to recognize and grasp novel concepts by learning from a limited number of known samples. The existing studies on few-shot learning focus only on recognizing novel categories while neglecting the understanding of base knowledge. This paper introduces the concept of local relationship learning and proposes a generalized representation method for local relationships to address the two major research problems in few-shot incremental learning tasks, i.e., inferior inter-class distinguishability and the difficult generalization of incremental categories. To enhance the distinguishability, this paper first adopts the local spatial relationship to regularize the incremental representation ability. To alleviate inductive biases caused by the lack of data in the incremental process, this paper proposes a spatial generalization prototype generation algorithm, which uses distribution characteristics to quickly generate virtual prototypes and promote the effective representation of samples. Benefiting from the meta-learning training mechanism, this paper proposes a joint locality and generalization awareness incremental learning framework, which effectively alleviates catastrophic forgetting and distinguishability difficulties by combining the local representation of the base category and the fast generalization constraint of the incremental category. Our experimental results demonstrated that the proposed method achieves state-of-the-art results on few-shot incremental learning tasks.