Incremental Image Classification Method Based on Integrated Transformer

ABSTRACT

In incremental learning scenarios, catastrophic forgetting remains a critical challenge when deploying neural networks under strict memory constraints and privacy‐preserving requirements. To address this issue, we propose a novel Transformer‐integrated framework (Boosting‐based Transformer Network (BT‐Net)) that effectively mitigates knowledge forgetting while maintaining classification performance. Our approach introduces two key components: one is a plug‐and‐play Transformer module that constructs a fusion dictionary bridging old and new feature representations, establishing inter‐task dependencies through attention mechanisms; the other is a SAMME‐inspired bootstrapping ensemble method that optimally combines predictions from both historical and current models. Extensive experiments conducted on CIFAR‐100 and ImageNet datasets for incremental learning demonstrate that the proposed method effectively alleviates catastrophic forgetting and improves model performance. The proposed method exhibits superior adaptability to real‐world scenarios with limited initial data.