计算机应用 ›› 2016, Vol. 36 ›› Issue (3): 795-799.DOI: 10.11772/j.issn.1001-9081.2016.03.795

• 虚拟现实与数字媒体 • 上一篇    下一篇

对比度与空间位置关系驱动的显著性检测

刘志远, 李华锋   

  1. 昆明理工大学 信息工程与自动化学院, 昆明 650500
  • 收稿日期:2015-08-24 修回日期:2015-10-21 出版日期:2016-03-10 发布日期:2016-03-17
  • 通讯作者: 李华锋
  • 作者简介:刘志远(1989-),男,安徽淮南人,硕士研究生,CCF会员,主要研究方向:计算机视觉、图像处理;李华锋(1983-),男,安徽阜阳人,副教授,博士,CCF会员,主要研究方向:模式识别、图像处理。
  • 基金资助:
    国家自然科学基金资助项目(61302041);云南省科技厅基础研究计划项目(2013FD011);昆明理工大学自然科学研究基金资助项目(KKZ3201303027)。

Saliency detection using contrast and spatial location-relation

LIU Zhiyuan, LI Huafeng   

  1. Faculty of Information Engineering and Automation, Kunming University of Science and Technology, Kunming Yunnan 650500, China
  • Received:2015-08-24 Revised:2015-10-21 Online:2016-03-10 Published:2016-03-17
  • Supported by:
    This work is partially supported by the National Natural Science Foundation of China (61302041), the Yunnan Applied Basic Research Projet(2013FD011) and the Natural Science Fund Project of Kunming University of Science and Technology(KKZ3201303027).

摘要: 针对已有方法不能很好地检测显著目标边界以及完整区域问题,提出一种基于超像素分割的图像显著性检测方法。首先,对原图像进行双边滤波降低局部颜色差异,使图像更加平滑、均匀,同时能够保留显著目标的边缘信息。然后通过计算局部窗口内像素的差异来实现显著目标边界的初步检测;滤波后的图像通过超像素分割将具有相同或相近颜色特征的像素划分到一个超像素区块内,在此基础上,综合考虑超像素区块的局部对比度与全局对比度以及空间分布关系来计算每个区块的显著值。最后,融合上述两部分的结果并通过引导滤波来对检测结果进行优化处理。在MSRA-1000公开数据集上与其他7种方法进行对比实验,所提方法的平均准确率为81.57%,平均召回率为77.13%,综合指标F-measure值为80.50%。实验结果表明,提出的方法能够很好地检测出显著目标边界与内部信息,均匀突出了显著区域,证明了所提方法的有效性和鲁棒性。

关键词: 显著性检测, 超像素分割, 对比度, 空间关系

Abstract: Concerning that the existing methods cannot well detect the salient object boundary and entire region, a new method based on super-pixel segmentation was proposed. Firstly, the bilateral filtering was employed on original images to reduce the local color difference and make the image smoother and more homogeneous; at the same time, the information of salient object edge was retained. The initial detection of salient object's edge was implemented by calculating the pixel' difference within the local window; super-pixel segmentation was adopted to filtered image so that the pixels with the same or similar color were divided into the same super-pixel block, based on this, the local contrast, global contrast and spatial distribution of super-pixel block were considered synthetically to calculate the salient value of each super-pixel block. Finally, the results of the above two parts were fused and optimized by guided filtering. The experiments were conducted on the international open data set MSRA-1000 compared with other seven methods. The average accuracy rate, average recall, and F-measure value of the proposed method are 81.57%, 77.13% and 80.50% respectively. The experimental results show that the proposed method can exact salient object in images effectively and robustly.

Key words: saliency detection, super-pixel segmentation, contrast, spatial relation

中图分类号: