基于自适应区域限制FCM的图像分割方法  

Adaptive Region Constrained FCM Algorithm for Image Segmentation

作　　者：李磊[1] 董卓莉[1] 张德贤[1] LI Lei;DONG Zhuo-li;ZHANG De-xian(College of Information Science and Engineering,Henan University of Technology,Zhengzhou,Henan 450001,China)

机构地区：[1]河南工业大学信息科学与工程学院,河南郑州450001

出　　处：《电子学报》2018年第6期1312-1318,共7页Acta Electronica Sinica

基　　金：河南省教育厅自然科学项目(No.15A520057);河南省科技厅自然科学项目(No.132102210494;No.162102210189);高层次人才基金(No.21476062);省属高校基本科研业务费专项资金(No.2016QNJH25)

摘　　要：提出一种基于自适应区域限制FCM(Fuzzy C-Means)的彩色图像分割方法,结合隐马尔科夫模型,把超像素具有区域一致性作为先验知识自适应融入到聚类过程中,以提升聚类性能.算法首先生成图像的超像素,计算像素对该超像素的贡献度,以此计算该超像素的区域隶属度函数;然后根据像素所属超像素是否具有主标签,选择像素级隶属度函数或区域级隶属度函数计算该像素的点对先验概率,以加强分割结果的区域一致性;其中,使用区域隶属度函数将引导聚类优化的方向,因此在迭代过程中去除未被使用的标签;最后迭代终止获得图像的分割结果.实验结果表明,相对于比较算法,本文算法的分割性能有显著提升.An image segmentation method based on robust regional constraint FCM（ Fuzzy C-Means） is proposed,which combines hidden Markov random filed（ HMRF） model with FCM. In order to improve the performance of the proposed method,the consistency of superpixels of the input image is adaptively used as a priori in clustering process. The proposed method first obtains the superpixels of the image,and for each superpixel,calculates a contribution of each pixel to the superpixel and the contributions are used to compute the superpixel＇s membership functions. And then the pointwise prior probabilities of pixels are calculated with pixel-level membership function or region-level membership function according to whether the superpixel to which the pixels belong has the dominant label. The use of region-level membership function is to guide the direction of clustering optimization,and thus there are some unused labels which are removed in the iteration process. Finally,the segmentation result is obtained after iteration stop. Experimental results demonstrate the good performance of the proposed method.

关 键 词：图像分割 模糊聚类 超像素 主标签 区域限制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]