Relational embeddings and class covariance measures for few-shot image classification
In the context of few-shot image classification,the structural and discriminative relationships between images have not been adequately emphasized,leading to imperfect feature embedding and unclear metric effects.To focus on the matching relationships between image features,a network structure based on relationship embedding and class covariance measurement is proposed.This is achieved by paying attention to the spatial information of in-dividual images and matching features between images,realizing the relationship of metric embedding features.The proposed structure utilizes self-correlation attention and interactional correlation attention to learn the feature rela-tionships within and between images.Specifically,the basic feature representation of individual images is mapped in-to self-correlation tensors,focusing on the spatial information of channels to learn structural information within im-age tensors.For image pairs,the mutual correlation information of the two image representations is calculated,gener-ating shared attention to obtain discriminative relational information between images.Finally,the Mahalanobis dis-tance is used to measure the feature relationships.After extensive experimental validation,the proposed method is e-valuated on three few-shot benchmark datasets and compared with various existing methods.The results demon-strate that the network structure exhibits a certain effectiveness in learning the relationships between features.
万方数据
维普
