Class feature Sub-space for few-shot classification

Few-shot learning is used in the development of models that can acquire novel class concepts from limited training samples, facilitating rapid adaptation to novel, intricate, and varied tasks encountered in real-world scenarios. Compared to meta-training, traditional batch training exhibits superior efficiency. However, when addressing few-shot learning tasks, the features extracted by batch-trained models often struggle to effectively represent novel class concepts, resulting in unsatisfactory performance. This challenge arises from two primary issues. First, the models lack the capacity to autonomously map novel class features into an effective discriminant subspace, resulting in the interference of class-independent feature components during metric calculations. Second, the limited sample size hinders the ability of the model to accumulate valuable experience, leading to representations that deviate from the true class center. To address these issues, we introduce the Class Feature Sub-space (CFS-space) as an effective discriminant space for few-shot classification; it preserves class features and suppresses noise components by mapping the extracted features into the CFS-space. Furthermore, we incorporate empirical knowledge from the base set and calibrate the prototypes within the CFS-Space to enhance the class representations. Ablation studies affirm the efficacy of our approach. As evaluated in 5-way 1/5-shot tasks, our method achieves impressive accuracies of 66.42%/83.69% on mini-ImageNet, 72.07%/86.36% on tiered-ImageNet, and 79.34%/90.42% on CUB, significantly narrowing the performance gap between the batch-training and the meta-training paradigm.