基于张量特征的小样本图像快速分类方法  

Tensor feature-based faster classification network for few-shot learning

作　　者：张艳莎 冯夫健 王杰[1,2] 潘凤 谭棉[1,2] 张再军 王林 Zhang Yansha;Feng Fujian;Wang Jie;Pan Feng;Tan Mian;Zhang Zaijun;Wang Lin(School of Data Science and Information Engineering,Guizhou Minzu University,Guiyang,550025,China;Key Laboratory of Pattern Recognition and Intelligent System,Guizhou Minzu University,Guiyang,550025,China;School of Mathematics and Statistics,Qiannan Normal University for Nationalities,Duyun,558000,China)

机构地区：[1]贵州民族大学数据科学与信息工程学院,贵阳550025 [2]贵州省模式识别与智能系统重点实验室,贵州民族大学,贵阳550025 [3]黔南民族师范学院数学与统计学院,都匀558000

出　　处：《南京大学学报（自然科学版）》2022年第6期1059-1069,共11页Journal of Nanjing University（Natural Science）

基　　金：国家自然科学基金(62162012);贵州省科技支撑计划(黔科合支撑[2021]一般531);贵州省教育厅自然科学研究项目(黔教技[2022]015号);贵州省教育厅深化新时代教育评价改革试点项目(教学过程质量评价);贵州省教育厅青年科技人才成长项目(黔教合KY字[2022]177号,黔教合KY字[2021]104号,黔教合KY字[2018]141号,黔教合KY字[2018]140,黔教合KY字[2021]110);贵州省科技计划(黔科合基础-ZK[2022]一般195,黔科合基础-ZK[2022]一般550)。

摘　　要：解决小样本图像分类问题最直接的方式是进行数据增强,但目前适用于小样本图像分类的数据增强方法大都存在模型复杂、推理时间长的问题.提出一个张量特征生成器,通过生成新的张量特征在特征空间对小样本图像进行数据增强.基于张量特征生成器,提出一个适用于小样本图像的快速分类方法(Tensor Feature-based Faster Classification Network,TFFCN),该方法网络结构简单,利用残差网络提取图像的张量特征,通过张量特征生成器对小样本图像进行数据增强,从而训练得到一个满意的分类器对查询集图像进行分类,解决了模型推理时间长的问题.为了验证提出模型的有效性,选用公开数据集miniImageNet,CUB以及CIFAR-FS,对分类性能和推理时间进行对比实验.实验结果表明,TFFCN的分类性能优于目前流行的数据增强方法,并且,和改进前的模型相比能有效减少模型的推理时间,采用ResNet18和ResNet12为主干特征提取网络时,随着生成的张量特征数量的增加,最高可减少49%和24%的推理时间,能更快速地完成小样本图像分类任务.The most straightforward way to solve the problem of few-shot image classification is data augmentation. Aiming at the problem that most of the data augmentation methods suitable for few-shot image classification are overly sophisticated and need a long inference time. We propose a tensor feature generator,which augments few-shot images in the features space by generating new tensor features. Based on the tensor feature generator,a rapid classification method for few-shot learning,Tensor Feature-based Faster Classification Network(TFFCN) is proposed. The network structure is simple,the tensor features of the image are extracted by using the residual network,and data augmentation of few-shot image with tensor feature generator,so as to train a satisfactory classifier and classify the query set images,and solve the problem of long inference time.We use public datasets miniImageNet,CUB and CIFAR-FS to verify the effectiveness of the proposed model by comparing experimental results on classification performance and inference time. Experimental results show that the classification performance of the TFFCN is better than popular data augmentation methods for few-shot image classification,and the inference time is greatly reduced compared to the model before the improvement. When the backbone networks are ResNet18 and ResNet12,the inference time is almost reduced by up to 49% and 24% with the increase of the number of generated tensor features,respectively,and the TFFCN can complete the task of few-shot image classification more quickly.

关 键 词：小样本图像分类 数据增强 张量特征生成器 张量特征 推理时间 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]