文献类型：期刊文献

中文题名：利用背景先验的显著性检测算法

英文题名：Saliency detection algorithm based on background prior

作者：张巧荣[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

第一机构：河南财经政法大学计算机与信息工程学院

年份：2016

卷号：21

期号：2

起止页码：165-173

中文期刊名：中国图象图形学报

外文期刊名：Journal of Image and Graphics

收录：CSTPCD;;北大核心:【北大核心2014】；CSCD:【CSCD2015_2016】；

基金：国家自然科学基金项目(61374079);河南省基础与前沿技术研究计划项目(122300410379);河南省教育厅科学技术重点研究项目(14A520025)~~

语种：中文

中文关键词：显著性检测;背景先验;多尺度分析;显著图;视觉显著性

外文关键词：saliency detection; background prior; multi-scale analysis; saliency map; visual saliency

摘要：目的针对图像的显著区域检测问题,提出一种利用背景先验知识和多尺度分析的显著性检测算法。方法首先,将原始图像在不同尺度下分解为超像素。然后,在每种尺度下根据各超像素之间的特征差异估计背景,提取背景区域,获取背景先验知识。根据背景先验计算各超像素的显著性,得到显著图。最后,将不同超像素尺度下的显著图进行融合得到最终显著图。结果在公开的MASR-1000、ECSSD、SED和SOD数据集上进行实验验证,并和目前流行的算法进行实验对比。本文算法的准确率、召回率、F-Measure以及平均绝对误差均在4个数据集上的平均值分别为0.718 9、0.699 9、0.708 6和0.042 3,均优于当前流行的算法。结论提出了一种新的显著性检测算法,对原始图像进行多尺度分析,利用背景先验计算视觉显著性。实验结果表明,本文算法能够完整、准确地检测显著性区域,适用于自然图像的显著性目标检测或目标分割应用。
Objective By focusing on the problem of visual saliency detection in images, an algorithm for saliency detection based on background prior and multi-scale analysis is proposed. Method The method consists of four steps. First, the orig- inal image is decomposed into super pixels. The sizes of salient regions vary; thus, the super pixel scale has a significant effect on the detection results. Therefore, the image is analyzed with different super pixel scales. Second, background re- gion is extracted. When extracting the background region, three rules are used, namely, boundary, connectivity, and fea- ture difference among the super pixels. The super pixels of the image are classified into background and foreground. Third, according to feature differences between the super pixel and background prior, the background-based saliency of the super pixels is calculated. Similarly, the foreground-based saliency can also be computed. The saliency map can be generated by integrating background-based saliency and foreground-based saliency. Finally, saliency maps under different scales are fused to obtain the final saliency map. Result To verify the efficiency of theproposed algorithm, we used four datasets, namely, MASR-1000, ECSSD, SED, and SOD datasets. The results are compared with state-of-the-art algorithms. We compare our algorithm and other state-of-the-art algorithms by four indicators ： precision, recall, f-measure, and mean abso-lute error （MAE）. Experimental results show that the proposed algorithm outperformsother current popular algorithms on MSRA-1000, SED, and SOD datasets. On the ECSSD dataset, our algorithm is similar to the manifold ranking algorithm. The average values of precision, recall, F-measure, and MAE are 0. 7189, 0. 6999, O. 7086, and 0. 0423, respectively. Conclusion In this paper, a novel saliency detection algorithm is proposed. According to the proposed algorithm, the origi- nal image is analyzed in muhi-scales. Visual saliency is computed by using the background prior. Experimental results show that the proposed algorithm can be successfully used in salient object detection and object segmentation in natural images.