基于矩形波谱分析的图像显著性检测  

Image saliency detection based on rectangular-wave spectrum analysis

作　　者：杨溢[1] 盛斌[1,2] 吴雯[3] 马利庄[1] 

机构地区：[1]上海交通大学计算机科学与工程系,上海200140 [2]中国科学院软件研究所计算机科学国家重点实验室,北京100190 [3]澳门大学科技学院电脑与资讯科学系,澳门999078

出　　处：《浙江大学学报（理学版）》2015年第1期1-8,共8页Journal of Zhejiang University（Science Edition）

基　　金：973计划资助(2011CB302203);国家自然科学基金资助项目(61202154;61133009);浙江大学CAD/CG国家重点实验室开放课题(A1401);中科院自动化所模式识别国家重点实验室开放基金项目;澳门大学科研基金项目(MYRG150(Y3-L2)/FST11/WW);上海市科委项目资助(13511505000);上海市浦江人才计划(13PJ1404500)

摘　　要：在计算机图形学与多媒体技术领域,显著性检测具有广泛的应用,已成为许多工作重要的一步.大量计算机图形学与多媒体技术工作(如图像分割、图像标记、视频跟踪)依赖于精确的显著性图.然而,现有的大多数显著性检测算法所得到的结果存在较大缺陷:如边界不明晰、召回率过低等.提出了一种基于矩形波谱分析的显著性检测算法,将图像分块,分别与预先设定好的矩形波构成的模板进行卷积运算,依据所得到的响应,评价某一区域与周围区域的差异性以及在全局中的独特性,赋予其相应的显著性值.通过在不同数据集上的大量实验证明,与现有的大部分显著性检测算法相比,本方法不仅有较高的准确率-召回率,且在得到的显著性图中,具有更明晰的边界.The saliency detection is an important step in many computer graphic and multi-media technology tasks be- cause of its wide applicability. Computer graphic tasks （such as image segmentation, image labeling and tracking） rely mainly on fine saliency maps. However, most of the existing methods are defective due to their poor ability in generating fine boundary or high recall rate. This paper proposes a saliency detection method. The given image is divided into several regions which convolute with a pre-set rectangle-wave template later. For each region, the final saliency value is determined by the difference between themselves and the regions next to them as well as their uniqueness when they are compared with the whole images according to the corresponding response. Repeated tests on the different data sets prove its high precision-recall rate. Additionally, the boundary in our saliency map is clearer and finer.

关 键 词：显著性检测 矩形波 谱分析 响应矩阵 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]