文献类型：期刊文献

中文题名：利用视觉显著性的前景目标分割

英文题名：Foreground object segmentation based on visual saliency

作者：张巧荣[1];徐国愚[1];张俊峰[1]

第一作者：张巧荣

机构：[1]河南财经政法大学计算机与信息工程学院,郑州450046

第一机构：河南财经政法大学计算机与信息工程学院

年份：2019

卷号：0

期号：6

起止页码：833-840

中文期刊名：兰州大学学报：自然科学版

收录：CSTPCD;;Scopus;北大核心:【北大核心2017】；CSCD:【CSCD2019_2020】；

基金：国家自然科学基金项目(60612153,31700858);河南省高等学校重点科研项目(18A520016);河南省科技攻关项目(182102210020).

语种：中文

中文关键词：目标分割;视觉显著性;显著图;背景先验

外文关键词：object segmentation;visual saliency;saliency map;background prior

摘要：针对图像中的前景目标分割问题,提出一种视觉显著性引导的前景目标分割算法.对原始图像进行预处理后分解为互不重叠的超像素区域.将这些区域构成一个无向图,相邻两个区域间存在边,通过计算相邻区域间的特征差异得到边的权值.提取图像边缘的超像素区域作为背景区域,利用无向图计算各超像素区域相对于背景区域的视觉显著性,得到初始显著图.对初始显著图进行改进和优化,根据视觉显著性计算结果采用自适应阈值进行前景目标分割.在公开的图像数据集MASR-1000、ECSSD、Pascal-S和SOD上进行实验验证,并和目前流行的算法进行对比.结果表明,本研究算法在查准率、召回率、平均绝对误差及F-Measure等方面优于目前流行的几种算法,用于图像和视频的前景目标检测与分割是正确有效的.
Focusing on the problem of foreground object segmentation in images and videos,an algorithm for foreground object segmentation based on visual saliency was proposed.The method consisted of three steps.The original image was preprocessed and decomposed into non-overlapping regions called super pixels.These regions were made up of an undirected graph.There were edges between two adjacent regions,and the weights of the edges were obtained by calculating the feature differences between the adjacent regions.The background regions were extracted from the edge regions of the image.The visual saliency of each region was calculated according to the feature differences between the region and the background regions based on the undirected graph.The saliency map was improved and optimized.Foreground objects were segmented based on the visual saliency results.To verify the efficiency of the proposed algorithm,it was evaluated on four static datasets,i.e.MASR-1000,ECSSD,Pascal-S and SOD.The results were compared with some state-of-the-art algorithms through four indicators:precision,recall,MAE and F-Measure.Experimental results showed that the proposed algorithm outperformed other current popular algorithms on these datasets and the proposed algorithm could segment the foreground objects completely and accurately.