Unsupervised descriptor selection based meta-learning networks for few-shot classification

Highlights：

• An unsupervised descriptor selection module is introduced to extract sample-related information in image, which removes irrelevant parts and promotes metric learning in adaptation.

• We propose a task-related feature aggregation module to enhance internal representations in meta-learning, which generates compact embeddings and further improves network adaptation ability.

• We conduct extensive experiments on datasets Caltech-UCSD Bird and miniImageNet. The experimental results demonstrate that with a simple structure, the proposed model obtains comparable performance and further promotes the classification accuracy when applied in prior meta-networks.