融合低层和高层特征图表示的图像显著性检测算法.pdfVIP

第 2 8 卷 第 3 期 计算机辅助设计与图形学学报 V ol. 28 No.3 2016 年 3 月 Journal of Computer-Aided Design Computer Graphics M ar. 2016 融合低层和高层特征图表示的图像显著性检测算法 高思晗 1〉，张 雷 1〉，李成龙 汤 进 1，2” 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) (tj@ ahu.edu.cn) 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1 0 种方法进行比较与分析，验证了该算法的有效性. 关 键 词 ：图像显著性；特征融合；图表示模型；不对称转移 中 图 法 分 类 号 ： TP391 Image Saliency Detection via Graph Representation with Fusing Low-level and High-level Features Gao Sihan 1), Zhang Le i1), L i Chenglong1)，and Tang Jin 1，2)* (School o f Computer Science and Technology, Anhui University, Hefei 230601) 2) (Key Laboratory o f Industrial Image Processing Analysis o f Anhui Province, Hefei 230039) Abstract : To employ complementary benefits o f 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 o f the nodes and edges are defined by high-level features and the difference o f low -level features, respec­ tively . Then, a sjmimetric transition probability m atrix 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 in itia l saliency map. To improve the robustness o f the proposed method, the center prior and the improved bound­ ary prior are integrated into our model. Extensive experiments on four publicly available datasets w ith ten approaches demonstrate the effectiveness o f the proposed approach. Key words : image sal