小样本图像语义分割综述  

A Survey on Few-Shot Image Semantic Segmentation

作　　者：陈琼[1] 杨咏 黄天林 冯媛 CHEN Qiong;YANG Yong;HUANG Tianlin;FENG Yuan(School of Computer Science and Engineering,South China University of Technology,Guangzhou,Guangdong 510006,China)

机构地区：[1]华南理工大学,计算机科学与工程学院,广东广州510006

出　　处：《数据与计算发展前沿》2021年第6期17-34,共18页Frontiers of Data & Computing

基　　金：广州市重点研发计划项目(202103010005);广东省国际合作项目(2021A0505030017);国家自然科学基金(62176095)。

摘　　要：【目的】对小样本语义分割方法进行系统而全面的介绍,为后续小样本分割算法设计工作提供参考。【方法】当前的小样本分割方法借助基于度量的元学习方法来完成少样本情况下的语义分割任务。根据度量工具是否可学习,将小样本分割算法分为基于参数结构和基于原型结构的小样本分割算法,简述了两类算法的优缺点。【结果】对该领域的一些经典工作和近年来的工作做了具体的分析,并给出了小样本分割算法的主要应用场景。【结论】在此基础上,分析了小样本分割存在的关键问题和挑战,对小样本分割未来的发展方向和趋势进行了讨论。[Objective]This paper introduces the few-shot image semantic segmentation methods systematically and comprehensively,as a reference to the design of the few-shot segmentation algorithm.[Methods]The metric-based meta-learning method is employed to perform the few-shot segmentation tasks.According to whether the metric tool is learnable,the few-shot segmentation algorithms are divided into prototype-based methods and parameter-based methods.This paper describes the advantages and disadvantages of both algorithms.[Results]Some classical and recent research works about the few-shot image semantic segmentation are analyzed in detail,together with an introduction of the main applications of the few-shot segmentation algorithm.[Conclusions]Hereinafter,the future development direction and trend of few-shot image segmentation are discussed,and its key problems and challenges are analyzed.

关 键 词：计算机视觉 图像数据 深度学习 图像分割 小样本学习 小样本语义分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]