Few-shot image classification aims to learn a model to correctly classify images with a few labeled data, however, feature extractors often fail to extract more generalized and discriminative features in low-data scenarios. Previous work has shown promising improvements by utilizing local descriptors of images. Unfortunately, they do not effectively suppress local descriptors where background noise is located and local descriptors that favour the classification of other classes. In this paper, we propose exploiting the relationship between the support set itself and the relationship between the support set and the query set in a task to enhance the discriminative power of local descriptors while suppressing the local descriptors associated with background noise in the images. In addition, we explore how to combine global features of images with local descriptors organically. Extensive experiments demonstrate that our method outperforms the state of the art on both standard datasets for few-shot learning.