关系嵌入和类协方差度量的小样本图像分类

Relational embeddings and class covariance measures for few-shot image classification

在小样本图像分类的问题情景下,图像之间具有结构性、辨别性的特征关系还未被重视,图像的特征关系匹配不足,导致特征嵌入不完善,度量效果不明显.为了关注图像特征之间的匹配关系,通过关注单个图像的空间信息及匹配图像间的特征,实现度量嵌入特征的关系,提出基于关系嵌入和类协方差度量的网络结构,通过自相关注意表示和交互相关注意来学习图像内和图像间的特征关系.其中,将单个图像的基本特征表示映射成自相关张量,关注其中的通道空间信息,学习图像张量内结构信息.而对于图像对,计算2个图像表示的互关联信息,并生成共同注意,从而得到图像间具有判别性的关系信息.最后用马氏距离度量特征关系.经过大量实验验证之后,在3个小样本基准数据集中评估,并与大量现有方法进行对比,结果显示网络结构对学习特征间关系具有一定的有效性.

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 individual 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 relationships within and between images.Specifically,the basic feature representation of individual images is mapped into self-correlation tensors,focusing on the spatial information of channels to learn structural information within image tensors.For image pairs,the mutual correlation information of the two image representations is calculated,generating shared attention to obtain discriminative relational information between images.Finally,the Mahalanobis distance is used to measure the feature relationships.After extensive experimental validation,the proposed method is evaluated on three few-shot benchmark datasets and compared with various existing methods.The results demonstrate that the network structure exhibits a certain effectiveness in learning the relationships between features.