融合颜色属性和空间信息的显著性物体检测  

Combining color names with spatial information for salient object detection

作　　者：徐丹[1,2] 唐振民[1] 徐威[1] 

机构地区：[1]南京理工大学计算机科学与工程学院,南京210094 [2]江苏科技大学计算机科学与工程学院,镇江212003

出　　处：《中国图象图形学报》2014年第4期541-548,共8页Journal of Image and Graphics

基　　金：国家自然科学基金重大研究计划重点项目(90820306)

摘　　要：目的提出了一种基于颜色属性和空间信息的显著性物体检测算法,并将其用于交通标志的检测。方法首先,训练颜色属性得到颜色-像素值分布,据此将图像划分为不同的颜色聚簇,每个聚簇的显著性取决于其空间紧致性。其次,将每个聚簇分割为多个区域,用颜色属性描述子表示每个区域,计算区域的全局对比度。最后,同时考虑区域对比度和相应聚簇的空间紧致性,得到最终的显著图。在此基础上,将交通标志的先验知识转换为自上而下的显著性图,形成任务驱动的显著性模型,对交通标志进行检测。结果在公开数据集上的测试结果表明,该算法取得最高92%的查准率,优于其他流行的显著性算法。在交通标志数据集上的检测取得了90.7%的正确率。结论提出了一种新的显著性检测算法,同时考虑了区域的颜色和空间信息,在公开数据集上取得了较高的查准率和查全率,在对交通标志的检测中也取得了较好的结果。Objective Visual saliency refers to the physical, bottom-up distinctness of image details. It is a relative property that depends on the degree to which a pixel or region is visually distinct from its background. The region-based contrast method （RC） achieves good results on public dataset. However, the over-segmentation of the method is considered as a preview contrast computation which contributes to the high precision. In this study, a novel bottom-up salient object detec- tion method based on color names and spatial information is proposed, in which regional contrast and spatial compactness are considered as two factors for saliency evaluation. In addition, we express the prior knowledge of traffic signs with top- down saliency maps, combined with the bottom-up saliency maps, to detect traffic signs from real world images. Method First, we learn color names offline by a probabilistic latent semantic analysis model to find the mapping between color names and pixel values. The color names can be used for image segmentation and region description. Second, pixels are as- signed into special color names according to their values to form different color clusters in the image. The saliency measure for every color cluster is evaluated by its intra spatial variance. The less the color cluster spreads the more salient it is. Third, every color cluster is divided into some local regions which are represented by color name descriptors. The region- al contrast is evaluated by computing color distance between different regions in entire image. Last, the final saliency map is constructed by incorporating the color cluster＇ s spatial compactness measure and the corresponding regional contrast. Note that in most cases, it is not the bottom-up saliency, but the most ＂interesting＇＇ object in an image that attracts attention. In this study, road signs are divided into three categories; every category has special color information. For each category, a class-specific distribution is constructed by the bag-of-words （

关 键 词：颜色属性 视觉显著性 区域对比度 空间紧致性 路标检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]