分层信息融合的物体级显著性检测

目的 显著性检测是基于对人类视觉的研究,用来帮助计算机传感器感知世界的重要研究手段。现有显著性检测方法大多仅能检测出人类感兴趣的显著点或区域,无法突出对象整体的显著性以及无法区分对象不同层次的显著性。针对上述问题,提出一种基于分层信息融合的物体级显著性检测方法。方法 与当前大多数方法不同,本文同时运用了中级别超像素和物体级别区域两种不同层次的结构信息来获取对象的显著图。首先,将图像分割为中级别的超像素,利用自下而上的方法构造初始显著图;然后通过谱聚类方法将中级别的超像素聚类成物体级的区域,并运用自上而下的先验来调整初始先验图;最后,通过热核扩散过程,将超像素级别上的显著性扩散到物体级的区域上,最终获得一致的均匀的物体级显著性图。结果 在MSRA1000标准数据库上与其他16种相关算法在准确率-召回率曲线及F度量等方面进行了定量比较,检测的平均精度和F-检验分数比其他算法高出5%以上。结论 通过多层次信息融合最终生成的显著图,实现了突出对象整体显著性以及区分不同对象显著性的目标。本文方法同样适用于多目标的显著性检测。

Object level saliency detection by hierarchical information fusion

Objective Saliency detection, which is based on the simulation of human visual attention, is an important way to help computer sensors to understand the world. Saliency detection has many applications in computer vision, such as image segmentation, image retrieval, and retargeting. However, saliency detection is a challenging computer vision task. Most of the existing saliency algorithms can only detect pixels or regions of interest. Method A new method based on hierarchical information fusion is proposed in this study to distinguish the saliency object from the complex background region and guarantee the uniformity of patches in the same object. The proposed method is different from the state-of-art method that uses mid-level superpixels and object-level regions to adjust the raw saliency map. First, an edge-preserving filtering is adopted as a pretreatment and then the mid-level superpixels are generated by the simple linear iterative clustering algorithm. Second, the mid-level raw saliency map is obtained by the saliency filter and adjusted by two priors, which can reduce the influence of complex background regions. Afterward, the mid-level superpixels are clustered to object-level segments by spectral clustering, and an object boundary prior is defined to enhance the consistency of the saliency map. Finally, the saliency label will be diffused from superpixels to object-level regions by heat diffusion. Result The evaluation experiments against 16 other methods are conducted on the benchmark MSRA1000 database by the precision-recall curve and the F-measure score. Conclusion By utilizing the mid-level superpixels and object-level clustering regions, our method can reflect the hierarchical relationship between patches and objects well. The experimental result show that our method is also applicable to multi-target saliency detection.