融合低层和高层特征图表示的图像显著性检测算法  

Image Saliency Detection via Graph Representation with Fusing Low-level and High-level Features

作　　者：高思晗 张雷[1] 李成龙[1] 汤进[1,2] 

机构地区：[1]安徽大学计算机科学与技术学院,合肥230601 [2]安徽省工业图像处理与分析重点实验室,合肥230039

出　　处：《计算机辅助设计与图形学学报》2016年第3期420-426,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家"八六三"高技术研究发展计划(2014AA015104);国家自然科学基金(61472002);安徽省高等学校省级自然科学研究项目(KJ2014A015);安徽省自然科学基金(1508085QF127)

摘　　要：为了有效地利用不同层次特征的互补性,提高鲁棒性,提出一种融合低层和高层特征的图表示的图像显著性算法.首先以超像素为结点构图,通过高层特征和底层特征差异定义该图的点和边的权重;然后根据该图模型构造不对称转移概率矩阵,并利用Markov随机游走算法进行求解,得到初始显著性图;最后结合中心先验及改进的边界先验得到最终的图像显著性结果.在4个公共数据集上与10种方法进行比较与分析,验证了该算法的有效性.To employ complementary benefits of different level features effectively and improve the robust-ness, we propose a graph representation based image saliency detection method, which fuses low-level and high-level features. We take superpixels as graph nodes to construct the graph model, in which the weights of the nodes and edges are defined by high-level features and the difference of low-level features, respec-tively. Then, a symmetric transition probability matrix is constructed based on the proposed graph represen-tation model, and the Markov random walk algorithm is utilized to optimize this model and obtain the initial saliency map. To improve the robustness of the proposed method, the center prior and the improved bound-ary prior are integrated into our model. Extensive experiments on four publicly available datasets with ten approaches demonstrate the effectiveness of the proposed approach.

关 键 词：图像显著性 特征融合 图表示模型 不对称转移 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]