电子学报 ›› 2018, Vol. 46 ›› Issue (7): 1683-1690.DOI: 10.3969/j.issn.0372-2112.2018.07.021

• 学术论文 • 上一篇    下一篇

联合Sobolev梯度的相场变分模型应用于图像显著性检测

李梦1,2, 刘星1, 任泽民3   

  1. 1. 重庆大学经济与工商管理学院, 重庆 400030;
    2. 重庆工商大学数学与统计学院, 重庆 400067;
    3. 重庆科技学院数理学院, 重庆 401331
  • 收稿日期:2017-02-13 修回日期:2017-08-19 出版日期:2018-07-25
    • 通讯作者:
    • 刘星
    • 作者简介:
    • 李梦,女,1973年出生,四川开江人,理学博士,现重庆工商大学副教授,研究方向:偏微分方程及其在图像处理中的应用.E-mail:limeng7319@email.ctbu.edu.cn;任泽民,男,1985年出生,山东枣庄人,理学博士,现重庆科技学院副教授,研究方向:偏微分方程及其在图像处理中的应用.E-mail:zeminren@cqu.edu.cn
    • 基金资助:
    • 国家自然科学基金 (No.61601068); 重庆市基础与前沿研究计划 (No.cstc2015jcyjA40029); 重庆市教委科学计划研究 (No.KJ1601317); 重庆工商大学校内科研 (No.2016-56-08,No.1752002)

A Phase Field Variational Model Joint Sobolev Gradient for Saliency Detection

LI Meng1,2, LIU Xing1, REN Ze-min3   

  1. 1. College of Economics and Business Administration, Chongqing University, Chongqing 400030, China;
    2. College of Mathematics and Statistics, Chongqing Technology and Business University, Chongqing 400067, China;
    3. College of Mathematics and Physics, Chongqing University of Science and Technology, Chongqing 401331, China
  • Received:2017-02-13 Revised:2017-08-19 Online:2018-07-25 Published:2018-07-25
    • Corresponding author:
    • LIU Xing
    • Supported by:
    • National Natural Science Foundation of China (No.61601068); Chongqing Research Program of Basic and Frontier Technology (No.cstc2015jcyjA40029); Scientific Research Project of Chongqing Municipal Education Commission (No.KJ1601317); Campus Scientific Research of Chongqing Technology and Business University (No.2016-56-08, No.1752002)

摘要: 基于变分框架和相变换理论,提出能量泛函应用于图像显著性检测,采用Sobolev梯度法最小化该能量泛函,由此导出基于时间的演化系统应用于计算机视觉选择,其检测过程是图像显著和非显著分量的竞争变化.比较传统的L2梯度,Sobolev度量具有更好的正则性.论文阐述了演化系统的动态行为和视觉选择之一致性,这对探索新的显著性检测机制非常重要.实验显示,本文模型在较好压制图像背景的同时,获得较为精细的目标、纹理,绒毛等信息,这些信息对于视觉感知是非常重要的.

关键词: 视觉注意, 显著性检测, 相场方法, 变分模型, Sobolev梯度

Abstract: An energy functional based on variational framework and the phase-transition theory is presented for saliency detection.We describe a Sobolev gradient to find minima of the proposed functional,and then a temporal evolution system for computer visual selection is produced.The process of saliency extraction is a dynamical competition between salient and non-salient component.Comparing the classical L2 gradient,we find that the Sobolev metric induces favorable regularity properties in their gradient flow.We demonstrate the consistency between the dynamical behavior of the evolution system and visual selection,which is very important for human to exploit new mechanism of saliency detection.Experimental results on various images show that our model achieves better suppression of the information in background,while achieving higher detection precision of the object,texture,hair,etc,which is important in terms of human visual perception.

Key words: visual attention, saliency detection, phase field method, variational model, Sobolev gradient

中图分类号: