基于符号规范化割的半监督学习与图像分割  

Semi-supervised Learning and Image Segmentation Base on Signed Normalized Cut

作　　者：刘先锋 石静 陈明 杨予丹 LIU Xian-feng;SHI Jing;CHEN Ming;YANG Yu-dan(College of Information Science and Engineering,Hunan Normal University,Changsha 410081,China)

机构地区：[1]湖南师范大学信息科学与工程学院,长沙410081

出　　处：《小型微型计算机系统》2020年第7期1547-1552,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(11871210)资助;湖南省自然科学基金项目(2018JJ3351)资助;湖南省教育厅科学研究项目(18C0016)资助。

摘　　要：半监督学习利用少量的辅助信息以提升学习器的性能.基于图的学习方法是较为典型的半监督学习实现途径,利用图来表达和分析数据,能够处理复杂的数据分布.不同于这类方法通常所利用的无符号图,符号网络具有更强的表达能力,其负边能够表达额外的信息.本文基于符号网络的规范化割(Signed Normalized Cut,SNCut),提出了可处理成对约束的半监督聚类,通过实验验证了负边给半监督学习带来的附加价值.将SNCut应用于图像分割问题,获得的分割效果明显优于规范化割.为了进一步强化边界对齐性,引入马尔科夫随机场(Markov Random Field,MRF)正则化项,构建SNCut&MRF目标函数,并提出基于界优化和图割的求解算法.结果表明,SNCut&MRF相比一些典型分割方法有更好的分割性能,在边界处表现良好.Semi-supervised learning uses a small amount of auxiliary information to improve performance.Graph-based learning is a typical way for semi-supervised learning,which employs graph to express and analyze data and applies to complex data distributions.Compared to unsigned graphs traditionally used by these methods,signed networks have stronger expressive abilities,and its negative links can express additional information.This paper proposes a semi-supervised learning method based on Signed Normalized Cut(SNCut in short)to deal with pairwise constraints and verifies additional value from negative links.SNCut is then applied to image segmentation,and significantly improved segmentation results of normalized cut.In order to further improve the boundary perfermance,a Markov Random Field(MRF in short)regularization term is introduced to construct a new objective function,called SNCut&M RF,which are then solved based on bound optimization and graph cuts.The experimental results show that SNCut&M RF has better segmentation performance than some classical segmentation methods and performs well at the boundary.

关 键 词：半监督学习 符号网络 规范化割 图像分割 马尔科夫随机场 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]