Joint Feature Disentanglement and Hallucination for Few-Shot Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Few-shot learning (FSL) aims to generalize from limited data.
• Current FSL methods often fail to fully utilize intra-class information for data hallucination.
• This can lead to hallucinated data that does not accurately represent novel categories.

Purpose of the Study:

• To propose a novel network, FDH-Net, for effective few-shot learning.
• To jointly perform feature disentanglement and data hallucination for improved FSL.
• To enhance the generalization capability of models to novel classes with limited examples.

Main Methods:

• FDH-Net disentangles visual data into class-specific and appearance-specific features.
• It utilizes both data recovery and classification constraints for hallucination.
• Appearance information from base categories is leveraged to hallucinate features for novel categories.

Main Results:

• FDH-Net demonstrates strong performance on both fine-grained (CUB, FLO) and coarse-grained (mini-ImageNet, CIFAR-100) datasets.
• The proposed framework outperforms existing state-of-the-art metric-learning and hallucination-based FSL models.
• Feature disentanglement and hallucination effectively improve generalization in few-shot scenarios.

Conclusions:

• FDH-Net offers a robust approach to few-shot learning by improving data hallucination.
• The method effectively leverages intra-class information through feature disentanglement.
• FDH-Net provides a significant advancement in generalizing to novel categories with minimal data.