中文摘要：

目前的视觉显著性检测算法,主要依赖像素间的对比,缺乏从全局角度对显著目标进行分析理解。根据生物视觉注意机制,显著目标通常是显眼、紧凑和完整的,提出一种基于混合图上随机游走的显著目标检测算法,将视觉显著性检测公式化为马尔科夫随机游走问题。首先将输入图像进行分块,利用颜色特征距离和方向的空间分布和方向熵对比分别确定无向图和有向图的边权重,进而得到混合图;然后通过全连通图搜索提取全局特性,突出全局较孤立的区域;同时通过k-regular图搜索提取局部特性,增强局部较均匀的区域;最后结合全局特性和局部特性得到输入图像的显著图,从而确定感兴趣区域位置。实验结果表明,相比于其他两种具有代表性的算法,所提算法检测结果更加准确、合理,证明该方法合理可行。

英文摘要：

Current visual saliency detection algorithms mainly focus on the inter-pixel contrast and lack the analysis of salient object from global perspective.According to biological visual attention mechanism that salient objects are compact and complete,a salient object detection algorithm with hybrid graph model is proposed,and the problem of salient region detection is formulated as Markov random walk model.First of all,the input image is divided into pixel blocks.A hybrid graph model is formed.The vertices are connected with undirected edges and directed ones.The undirected edges represent the color difference between two block images.The directed edges represent the dependence of the spatial distribution and regional complexity of the orientation feature.Secondly,the isolated regions are obtained using the random walk on a complete graph to extract the global properties.Meanwhile,the local uniform regions are enhanced using the random walk on a k-regular graph to extract the local properties.Finally,the saliency map is obtained through combining the global properties and local properties.The salient object is located according to the saliency map.Experimental results show that the proposed method is more reasonable than other two representative algorithms.