Integrating Task Information into Few-Shot Classifier by Channel Attention

It has been increasingly recognized that meta-learning-based approaches provide a promising way to handle challenges to few-shot learning. In this paper, we incorporate the channel attention in the main framework of simple-CNAPS proposed by Bateni et al. to develop a model more appropriate for few-shot image classification. In detail, we replace FiLM layers in simple-CNAPS with channel attention blocks which scale the image channels according to the relationship between task information and feature maps rather than only the task information. This replacement makes the feature extractor more expressive. Moreover, it allows us to take the interaction of different image channels into account. In addition, to alleviate the computational bias caused by small sample size, we provide a method to estimate class centers with perturbations. Finally, the effectiveness of the model is verified by experiments on the few-shot image classification benchmark datasets.