Optimising Few-Shot Class-Incremental Learning for Fine-Grained Visual Recognition

Traditional Deep Learning models often rely on extensive training datasets and struggle in dynamic environments. Few-Shot Class-Incremental Learning (FSCIL) addresses these challenges by enabling models to learn new classes with limited samples while retaining old knowledge. However, FSCIL models typically perform well on coarse-grained datasets but falter on fine-grained ones due to high inter-class similarity and limited samples for new classes. This paper introduces a novel framework that mitigates catastrophic forgetting by leveraging the generalization capabilities of the CLIP image encoder through Cosine-based Distillation. Additionally, we employ supervised contrastive learning and a forward-compatible strategy to address the challenges specific to Fine-Grained FSCIL (FG-FSCIL). Our method demonstrates state-of-the-art performance on four fine-grained datasets (CUB200, Stanford Dogs, Stanford Cars, and Aircraft), achieving significant improvements over existing methods. The code and datasets are available at https://github.com/yahuuuuuuu/Fine-Grained-FSCIL .