长尾识别研究进展  

Research Advance in Long-tailed Recognition

作　　者：张明[1,2] 翟俊海[1,2] 许垒 高光远 Zhang Ming;Zhai Junhai;Xu Lei;Gao Guangyuan(School of Mathematics and Information Science,Hebei University,Baoding 071002,China;Hebei Key Laboratory of Machine Learning and Computational Intelligence,Hebei University,Baoding 071002,China)

机构地区：[1]河北大学数学与信息科学学院,河北保定071002 [2]河北大学河北省机器学习与计算智能重点实验室,河北保定071002

出　　处：《南京师范大学学报（工程技术版）》2022年第2期63-72,共10页Journal of Nanjing Normal University(Engineering and Technology Edition)

基　　金：河北省科技计划重点研发项目(19210310D);河北省自然科学基金项目(F2021201020).

摘　　要：长尾识别是目前深度学习领域最热门的研究方向之一,长尾识别的工作重点是解决长尾分布数据的计算机视觉识别任务.长尾分布的显著特征为2-8分布,即20%的类占据80%的样本.将少数几个类占据了大部分数据的类称之为头部类;而大多数类占据了很少部分数据的类称之为尾部类.首先,列举解决长尾识别问题的各种方法.然后,将其划分为重采样、重加权、迁移学习、解耦特征学习和分类器学习以及其他方法进行阐述.最后,阐述对相关方法的理解.Long tail recognition is one of the most popular research directions in the field of deep learning.The focus of long tail recognition is to solve the computer vision recognition task of long-tail distributed data.The prominent feature of the long-tail distribution is the 2-8 distribution,that is,20%of the classes account for 80%of the sample.We call a class with a few classes that make up most of the data a header class.Classes where most classes occupy a small portion of the data are called tail classes.Firstly,various methods are introduced to solve the problem of long tail recognition.Then,they are divided into resampling,re-weighting,transfer learning,decoupling feature learning,classifier learning and other methods.Finally,our understanding of the related methods are introduced.

关 键 词：深度学习 长尾识别 计算机视觉 研究方法 神经网络 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]