基于对比度与最小凸包的显著性区域检测算法  

Salient region detection algorithm based on contrast and minimum convex hull

作　　者：范敏[1] 陈曦[1] 王楷[1] 李志勇[2] 王晓峰[2] 

机构地区：[1]重庆大学自动化学院,重庆400030 [2]国网重庆市电力公司江北供电分公司,重庆401147

出　　处：《仪器仪表学报》2014年第10期2328-2334,共7页Chinese Journal of Scientific Instrument

基　　金：中央高校基本科研业务费专项基金(CDJZR10170009);重庆市科技攻关项目(CSTC2012GG-YYJS40008);国家电网公司科技(SGCQJB00FZJS1400341)资助项目

摘　　要：显著性检测算法常通过计算像素或像素块之间的对比度来确定显著性,但是图像背景中经常会出现特殊区域与图像其他部分也有较大的对比度,导致基于对比度的显著性检测算法无法将这部分背景区域与主要目标区分开。提出一种基于对比度与最小凸包的显著性区域检测算法。以超像素作为基本计算单位,使用Wasserstein距离衡量超像素之间的差异,通过计算超像素间的全局与局部对比度得到对比度显著图;找出图像中特征点Harris角点的最小凸包,以最小凸包几何中心为中心点,根据每个超像素与中心点的距离计算中心显著图;最后将对比度显著图与中心显著图相结合得到最终的显著图,这种算法可以有效地将背景中具有高对比度的区域区分开。在Corel和MSRA图像数据库上进行仿真实验,结果表明该文所提算法对显著区域检测的查准率、查全率等仿真评价指标相对于传统算法都有明显的提升。Current salient region detection algorithms are mainly based on the contrast of pixels or pixel blocks. But in many images, some special parts of background also have high contrast, and many algorithms based on contrast cant distinguish these parts of the back- ground from the main targets. In order to solve this problem, this paper presents a new method for salient region detection based on con- trast and minimum convex hull. The method takes superpixel as the basic processing unit and measures the difference between superpix- els with Wasserstein distance; then the contrast saliency map is obtained through calculating the global and local contrasts between super- pixels. Next, this algorithm calculates the minimum convex hull of Harris corner points, and uses the central point of the minimum con- vex hull as the image central point. Based on the distances between each superpixel and the image central point, the center saliency map can be achieved. The final saliency map consists of the contrast saliency map and the center saliency map. This algorithm can effectively distinguish the region with high contrast in the background. Simulation experiments on the Corel and MSRA image databases were con- ducted; and the results show that compared with four representative algorithms, this salient region detection algorithm performs better in simulation evaluation indexes, such as precision, recall and so on.

关 键 词：显著性区域检测 超像素 Wasserstein距离 HARRIS角点 最小凸包 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]