一种前景和背景提取相结合的图像显著性检测  

Saliency Detection Based on Foreground and Background Extraction

作　　者：周强强[1] 赵卫东[1] 柳先辉[1] 王志成[1] Zhou Qiangqiang;Zhao Weidong;Liu Xianhui;Wang Zhicheng(CAD Research Center, College of Electronics & Information Engineering, Tongji University, Shanghai 201804)

机构地区：[1]同济大学电子与信息工程学院CAD研究中心,上海201804

出　　处：《计算机辅助设计与图形学学报》2017年第8期1396-1407,共12页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家科技支撑计划项目(2015BAF17B00)

摘　　要：为了获得更加精细化的显著目标检测结果,提出一种结合前景和背景信息的图像显著目标检测算法,将自底向上的粗糙显著区域提取和基于流形查询的自顶向下背景权重图的计算整合到统一的优化框架内.粗糙显著图主要融合了更符合生物心理学规则的局部对比图、频率先验图和全局颜色分布图这3个先验图;在背景权重图的计算中,首先根据超像素分割图构建一个无向图的邻接矩阵,然后基于边界背景先验知识选择位于图像边界的一些超像素作为初始流形查询向量进行图节点间关联度的传播计算,得到背景权重图.在MSRA1000和ECSSD这2个基准数据集上与当前主要的10种算法进行了对比实验,结果体现了文中算法的优异性.To obtain a more refined and accurate result of a salient detection object,in this paper we proposeda novel salient object detection algorithm which takes both background and foreground cues into consideration,and this algorithm integrates a bottom-up coarse salient regions extraction and a top-down backgroundweight map measure into a unified optimization framework.Where in the coarse saliency map isfused by three prior components,the first is local contrast map which is more accordance with the biopsychologylaw,the second is frequency prior map,and the third is the color distributed prior map.During thecomputation of the background weight map,we first construct an undirected graph based on superpixelsegmentation and select nodes on the border as an initial query to represent the background,then we performa relevance propagation to generate the background weight map.Comprehensive comparisons with10state-of-the-art solutions on two benchmark datasets-MSRA1000and ECSSD indicate that our algorithmwith a superior performance.

关 键 词：显著性检测 前景 背景权重图 流行查询 优化 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]