小样本图像分类的注意力全关系网络  

Total Relation Network with Attention for Few-Shot Image Classification

作　　者：李晓旭[1] 刘忠源 武继杰 曹洁[1] 马占宇[2] LI Xiao-Xu;LIU Zhong-Yuan;WU Ji-Jie;CAO Jie;MA Zhan-Yu(School of Computer and Communication,Lanzhou University of Technology,Lanzhou 730050;Pattern Recognition and Intelligent System Laboratory,School of Artificial Intelligence,Beijing University of Posts and Telecommunications,Beijing 100876)

机构地区：[1]兰州理工大学计算机与通信学院,兰州730050 [2]北京邮电大学人工智能学院模式识别与智能系统实验室,北京100876

出　　处：《计算机学报》2023年第2期371-384,共14页Chinese Journal of Computers

基　　金：国家自然科学基金(62176110,62111530146,61906080,61922015,U19B2036,62225601);甘肃省青年博士基金(2021QB-038);北京市自然科学基金(Z200002);兰州理工大学红柳杰出青年基金资助.

摘　　要：传统的基于深度学习的图像分类方法在大样本分类任务中具有较好的分类效果,但在小样本分类任务中却存在较大的挑战,为此,小样本图像分类获得了研究人员的广泛关注.基于度量的方法是解决小样本图像分类的一种简单有效方法,它利用可学习的映射函数将分类任务中的所有样本映射到一个特征空间中,然后基于某种度量标准对查询特征进行分类.由于分类任务中不同类的两个图像有可能包含较多的相似性区域,导致特征空间中某些查询特征与异类的类原型特征的距离较近,较难学习到大的分类边界.为了解决上述问题,本文提出了注意力全关系网络(Total Relation Network with Attention,TRNA),该网络通过计算特征对的全关系和特征对的注意力来实现大边界的特征空间.具体地,在计算出所有的查询特征和类原型后,提出的网络利用特征对全关系拼接操作将特征空间中的任意两个特征在通道方向上进行拼接得到特征对矩阵,然后利用特征对注意力机制将特征对矩阵中不同类间难区分的特征对挑选出来并给予大的权重,最后将特征对矩阵输入卷积网络和全连接网络得到一个相似得分矩阵.实验结果表明本文的方法与关系网络相比,在数据集mini-ImageNet、Stanford-Dogs、Stanford-Cars、CUB-200-2011的1-shot和5-shot分类任务中分别有2.67%和1.71%、8.31%和3.92%、14.99%和8.00%、4.41%和4.42%的性能提升.With the continuous development of deep learning,image classification methods based on deep learning have achieved excellent classification performance in large sample classification tasks,but face significant challenges in few-shot classification tasks.It is difficult to obtain a large number of labeled samples to train deep learning models in many real-world scenarios.This means that it is of great practical importance to improve the classification performance of deep learning-based image classification methods in few-shot classification tasks.To this end,a growing number of researchers are focusing on few-shot image classification,which aims to complete the classification of unlabeled query samples based on a small number of labeled support samples,that is,to learn new category concepts through a small number of labeled samples.The metric-based method is simple yet effective method for solving the few-shot image classification.It uses the learnable mapping function to map all samples in a few-shot classification task to feature space and then classifies the query features according to some metric standard.However,two images of different classes in the classification task may contain more similar regions,so there may be situations in the feature space where the distances of certain query features and heterogeneous class prototypes are closer to each other,resulting in few-shot image classification networks that are more difficult to learn large classification margin.In other words,it is more difficult to have a clear classification margin between feature clusters of different classes and significant compactness between features in feature clusters of the same class in the feature space.To solve this problem,this paper proposes a Total Relation Network with Attention(TRNA),which realizes the feature space with a large margin by calculating the Total Relation of Feature-Pair and the Attention of Feature-Pair.Specifically,after calculating all the query features and class prototypes,the proposed network uses the T

关 键 词：小样本图像分类 基于度量的方法 类原型 注意力机制 大边界学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]