利用类内类间信息的原型补足小样本图像分类  

Prototype⁃complemented few⁃shot image classification with intra⁃class and cross⁃class information

作　　者：吕佳[1,2,3] 郑小琪 LüJia;Zheng Xiaoqi(College of Computer and Information Sciences,Chongqing Normal University,Chongqing,401331,China;Chongqing Digital Agriculture Service Engineering Technology Research Center,Chongqing 401331,China;National Center for Applied Mathematics in Chongqing,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学计算机与信息科学学院,重庆401331 [2]重庆市数字农业服务工程技术研究中心,重庆401331 [3]重庆国家应用数学中心,重庆401331

出　　处：《南京大学学报（自然科学版）》2024年第4期600-612,共13页Journal of Nanjing University（Natural Science）

基　　金：国家自然科学基金(11991024);重庆市教委“成渝地区双城经济圈建设”科技创新项目(KJCX2020024);重庆市教委科研项目(KJZD-K202200511);重庆市高校创新研究群体资助项目(CXQT20015)。

摘　　要：基于度量学习的小样本学习方法中模型没有充分挖掘类内样本与类间样本的联系,将单个样本特征视作独立特征用作训练,导致模型生成的原型不准确且特征表示能力差.提出利用类内类间信息的原型补足小样本图像分类模型.首先,将支持集样本的特征送入类内信息提取分支,构造类内信息特征图,并提取特征获取类别描述信息为原型进行补足;然后,利用类间信息提取分支将不同类的查询样本融合生成新样本,并将该组查询样本的标签作为软标签;最后,使用补足后的原型对查询样本与新样本进行分类,并通过分类损失优化模型.在四个公开数据集上的实验结果表明,在MiniImageNet数据集上,准确率提升2.03%~5.48%;在TieredImageNet数据集上,准确率提升2.25%~8.55%;在CUB数据集上,准确率提升5.10%~8.82%;在CIFAR-FS数据集上,准确率提升2.61%~10.03%.证明与同类小样本图像分类方法相比,提出的模型获得了更优的分类性能与泛化性.The metric⁃based few⁃shot learning models struggle to fully exploit the relationship between the intra⁃class samples and the cross⁃class ones,and treat a single sample feature as an independent item during training,which results in inaccurate prototypes and low⁃quality representation.To handle the issue,a prototype⁃complemented few⁃shot image classification with intra⁃class and cross⁃class information is proposed.Firstly,the features of the support set are fed to the intra⁃class information extraction branch to exploit the intra⁃class feature,which is further processed to obtain the information of the category description to complement the initial prototypes.Then,the query samples from different classes are fused to generate the new samples through the cross⁃class information extraction branch.The labels of the query samples are constructed as soft labels of the new samples.Finally,the complemented prototypes are employed to classify the query and new samples,and the model is optimized by classification loss.In this paper,the comparison experiments are performed on four public few⁃shot learning datasets.The accuracy is improved by 2.03%to 5.48%,2.25%to 8.55%,2.61%to 10.03%,and 5.10%to 8.82%on the MiniImageNet dataset,TieredImageNet dataset,CUB dataset and CIFAR⁃FS dataset,respectively.Experimental results show that the proposed model achieves superior generalization and classification performance than other methods.

关 键 词：小样本学习 度量学习 元学习 类内类间信息 原型补足 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]