带噪声标签细粒度图像分类方法研究综述  

Review on Fine-Grained Image Classification with Noise Labels

作　　者：王浩 姚俊萍 李晓军 阙小翔 WANG Hao;YAO Junping;LI Xiaojun;QUE Xiaoxiang(Rocket Force University of Engineering,Xi’an 710025,Shaanxi)

机构地区：[1]火箭军工程大学,陕西西安710025

出　　处：《火箭军工程大学学报》2024年第5期96-108,共13页Journal of Rocket Force University of Engineering

基　　金：国家社科基金(2023-SKJJ-B-063)。

摘　　要：要:开展带噪声标签细粒度图像分类方法综述,在细粒度图像分类问题定义及方法类型梳理的基础上,对噪声标签问题进行描述,总结了标签噪声的来源,从影响因素依赖关系和实例依赖关系两个角度对标签噪声进行分类,分析了标签噪声对细粒度图像分类任务的影响,从基于显式噪声模型方法、基于隐式噪声模型方法两个维度梳理了细粒度图像分类的研究进展,提出了未来值得关注的研究主题。A review of classification methods of fine-grained images with noise labels was conducted.Firstly,based on the definition of fine-grained image classification problems and the classification of method types,the problems of noise labels were described,as well as the sourc-es of label noise were summarized.Then,label noise was classified from the perspectives of in-fluence factor dependency and instance dependency,and the impact of label noise on fine-grained image classification was analyzed.Furthermore,the progress of fine-grained image classification was discussed from two dimensions:explicit noise model based methods and implicit noise model based methods.Thereby,future noteworthy research topics were proposed.

关 键 词：细粒度图像分类 噪声标签 深度学习 噪声鲁棒 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]