电子学报 ›› 2020, Vol. 48 ›› Issue (8): 1509-1515.DOI: 10.3969/j.issn.0372-2112.2020.08.008

• 学术论文 • 上一篇    下一篇

超像素内容感知先验的多尺度贝叶斯显著性检测方法

张荣国1, 贾玉闪1, 胡静1, 刘小君2, 李晓明1   

  1. 1. 太原科技大学计算机科学与技术学院, 山西太原 030024;
    2. 合肥工业大学机械工程学院, 安徽合肥 230009
  • 收稿日期:2019-07-29 修回日期:2020-01-20 出版日期:2020-08-25 发布日期:2020-08-25
  • 作者简介:张荣国 男,1964年10月出生,山西太原人,合肥工业大学工学博士.现为太原科技大学计算机学院教授、硕士生导师,主要研究方向为计算机视觉、图像处理与模式识别. E-mail:rg_zh@163.com;贾玉闪 女,1993年5月出生,河北石家庄人,现为太原科技大学硕士研究生,研究方向为视觉显著性. E-mail:809132498@qq.com
  • 基金资助:
    国家自然科学基金(No.51875152);山西省自然科学基金(No.201801D121134);晋城市科技局资助项目(No.201501004-5)

Superpixel Content-Aware Priors Based Multi-Scale Bayesian Saliency Detection

ZHANG Rong-guo1, JIA Yu-shan1, HU Jing1, LIU Xiao-jun2, LI Xiao-ming1   

  1. 1. School of Computer Science and Technology, Taiyuan University of Science and Technology, Taiyuan, Shanxi 030024, China;
    2. School of Mechanical Engineering, Hefei University of Technology, Hefei, Anhui 230009, China
  • Received:2019-07-29 Revised:2020-01-20 Online:2020-08-25 Published:2020-08-25

摘要: 针对复杂背景下显著性检测方法不能够有效地抑制背景,进而准确地检测目标这一问题,提出了超像素内容感知先验的多尺度贝叶斯显著性检测方法.首先,将目标图像分割为多尺度的超像素图,在每个尺度上引入内容感知的对比度先验、中心位置先验、边界连通背景先验来计算单一尺度上的目标显著值;其次,融合多个尺度的内容感知先验显著值生成一个粗略的显著图;然后,将粗略显著图值作为先验概率,根据颜色直方图和凸包中心先验计算观测似然概率,再使用多尺度贝叶斯模型来获取最终显著目标;最后,使用了3个公开的数据集、5种评估指标、7种现有的方法进行对比实验,结果表明本文方法在显著性目标检测方面具有更好的表现.

关键词: 显著性, 多尺度, 内容感知先验, 边界连通性, 贝叶斯模型

Abstract: Existing saliency detection methods can not suppress the background effectively and detect the salient object accurately in complex background,a method of superpixel content-aware priors based multi-scale Bayesian saliency detection is proposed.Firstly,the image containing object is segmented into multi-scale superpixel maps,then the content-aware priors of contrast priors,center position priors,and boundary connected background priors are introduced on each scale to calculate the salient object values on a single scale;Secondly,the content-aware priors values of the various scales generate a rough saliency map;Thirdly,the rough saliency map value is used as the prior probability,and the likelihood is calculated according to the color histogram and the convex hull center,using the multi-scale Bayesian model to obtain the final salient object ;Finally,three public data sets,five evaluation indicators,and seven existing methods are used for comparative experiments.The experiments show that the method has better performance in the detection of salient objects.

Key words: saliency, multi-scale, content-aware prior, boundary connectivity, Bayesian model

中图分类号: