基于图像理解视角的分割全局评价算法  

Image Understanding Based Global Evaluation Algorithm for Segmentation

作　　者：钱晓华[1] 郭树旭[1] 李雪妍[1] 

机构地区：[1]吉林大学电子科学与工程学院,吉林长春130012

出　　处：《电子学报》2012年第10期1989-1995,共7页Acta Electronica Sinica

摘　　要：提出了一种具有全局属性的图像分割质量客观评价算法.该方法是基于实际分割结果与参考分割之间的差异进行计算的.从图像理解的角度出发,根据空间信息给予误分像素不同的权重,使得不同位置的误分像素具有不同的重要性;在加权公式中引入了相对距离,前景和背景搜索距离,以获得评价算法对于分割图像具有缩放不变性的结论.为了克服大比例过/欠分割造成的分割质量评价失真的现象,提出一种失真惩罚项,提高了评价算法的有效性和全局性.实验结果表明,本文提出的分割评价算法不仅能体现图像理解等后续处理的属性,而且具有更强的全局性和鲁棒性.An objective global evaluation algorithm of segmentation results is proposed, which is based on computing the de viation of the segmentation results from reference segmentation. By taking into account image understanding view, the discrepancy between two results is weighted based on spatial contextual information,thereby obtaining the different weightiness.Moreover,rela tive distance, foreground and background searching distance are introduced into the proposed rnelric, and the value of this melric is invariant to the segmentation images with varying scale. In order to overcome the in-afional evaluation caused by the excessive over- segmentation and under-segmentation, a penalized term is proposed. Experiment results prove that our evaluation accords with the view of the image understanding and it is more global and robust.

关 键 词：分割评价 差异法 理解权重 缩放不变性 惩罚项 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]