Few-Shot Class-Incremental Learning with Class Centers and Contrastive Learning for Incremental Vehicle Recognition
Huiyu Yi · 2024
The aim of few-shot class-incremental learning (FSCIL) is to continually classify new classes from limited samples without forgetting previously learned classes. Existing deep learning-based methods for fine-grained vehicle recognition are mostly deployed after a single model training session on a complete, large dataset. However, the real-world scenario involves constant emergence of new vehicle categories with limited labeled data. Retraining models at scale for every new vehicle release becomes impractical and this poses a FSCIL challenge. To address this issue, this paper proposes the Contrastive Center Optimization (CCO) method. Based on an Incremental-frozen framework, CCO prevents the forgetting of previous learning during the incremental sessions. In the base training session, CCO uses contrastive learning after data augmentation to indirectly decrease intra-class distances and increase inter-class distances. Additionally, we apply center loss to further directly optimize intra-class and inter-class distances. This approach not only enhances classification accuracy, but also creates sufficient space for the insertion of new classes. Extensive experiments on two fine-grained vehicle datasets demonstrate that our CCO achieves new state-of-the-art performance, confirming the importance of decreasing intra-class distances and increasing inter-class distances.