关键词: 关键词: 类增量学习, 深度学习, 数据增强, 正交约束, 灾难性遗忘

Abstract: Abstract: In class-incremental learning, overcoming catastrophic forgetting is a core challenge. Replay-based methods use a small memory buffer to store and replay old samples to mitigate forgetting. The scarcity number of old class samples can lead to overfitting, reducing the model's generalization ability. An Orthogonal Constraint and Attention-Guided Augmentation method (OCAGA) was proposed. An Attention-Guided Data Augmentation mechanism (AGDA) was designed to enhance the diversity of old class samples, alleviating the imbalance between new and old classes. An Orthogonal Projection Loss (OPL) was introduced, which explicitly utilized class information to pull samples of the same class closer and push samples of different classes apart in the embedding space, thus enhancing feature discriminability. Experiments on the CIFAR-100 and ImageNet-100 datasets demonstrate that OCAGA outperforms compared with incremental Classifier and Representation Learning (iCaRL), achieving an increase in final accuracy of 2.55 and 4.12 percentage points, respectively, thereby validating its effectiveness in mitigating catastrophic forgetting and promoting knowledge retention. 

Key words: class-incremental learning, deep learning, data augmentation, orthogonal constraint, catastrophic forgetting