摘要
相似度学习方法通过学习合适的相似度度量以改进模型的分类或聚类效果。现有的研究表明,相似度学习方法在很多计算机视觉问题中起到重要的作用。近年来随着数据规模的急剧增大和应用领域的多样化,相似度学习问题发展了很多新的研究领域。本文介绍了近年来相似度学习问题的研究进展和发展过程,包括从传统的二元组和三元组约束发展到新型相似度约束、从欧氏距离与马氏距离发展到新型相似度度量、从图像间的相似度学习发展到图像集之间的相似度学习、从单一模态相似度学习发展到跨模态相似度学习。最后本文展望了相似度学习未来可能的发展方向。
The similarity learning method aims to improve the classification or clustering performance by learning the appropriate similarity metric.The existing researches demonstrate that the similarity learning methods play very important roles in a number of computer vision applications.With the rapid growing data scale and the variant application fields in recent years,the similarity learning problem develops many new research fields.This paper introduces the research development progresses of similarity learning in recent years,i.e.from the traditional doublet and triplet constraints to new similarity constraint,from Euclidean distance and Mahalanobis distance to new similarity measure,from image based similarity learning to image set based similarity learning,from single modality similarity learning to multi-modality similarity learning.Finally,this paper prospects the possible develop direction of similarity learning in the future.
作者
王法强
张宏志
王鹏
邓红
张大鹏
WANG Faqiang;ZHANG Hongzhi;WANG Peng;DENG Hong;ZHANG Dapeng(School of Computer Science and Technology,Harbin Institute of Technology,Harbin 150001,China;College of Science,Northeast Agricultural University,Harbin 150030,China)
出处
《智能计算机与应用》
2019年第1期149-152,158,共5页
Intelligent Computer and Applications
基金
国家自然科学基金(61671182
61471146)
关键词
相似度学习
距离度量学习
深度相似度学习
similarity learning
distance metric learning
deep similarity learning
作者简介
王法强(1989-),男,博士研究生,主要研究方向:计算机视觉、相似度学习。
