一种局部受限的规则聚类超像素算法  

Locally-restricted regular clustering superpixel algorithm

作　　者：汪云飞[1] 毕笃彦[1] 刘华伟[1] 刘凌[1] 赵晓林[1] 

机构地区：[1]空军工程大学航空航天工程学院,陕西西安710038

出　　处：《西安电子科技大学学报》2016年第3期95-100,共6页Journal of Xidian University

基　　金：国家自然科学基金资助项目(61203268;61379104)

摘　　要：传统的超像素算法复杂度高,难以得到紧致且边缘贴合度好的超像素,针对此问题提出一种局部受限的规则聚类超像素算法.以k均值算法为基础,采用局部受限的聚类方法和类合并策略,得到大小均衡、外形规整的超像素.在聚类时充分考虑了像素点的颜色和位置特征,引入对数机制平衡两者在数值上的差异性,并通过特殊的颜色距离滤波使超像素的边缘更为光滑.仿真实验表明:所提算法简单易用,计算效率高,能够得到边缘重合率高且欠分割错误率低的超像素.当分割的超像素数较多时,性能要优于其他几种优异的超像素算法.It is difficult to obtain superpixels which are compact and adhere well to image boundary by traditional superpixel algorithms, because of their high complexity. This research proposes a new superpixel algorirhm of Locally-Restricted Regular Clustering （LRRC） for overcoming those difficulties. This algorithm is based on the k-means algorithm, and adopts the LRRC method and a class combination strategy to produce superpixels with equal and regular sizes. In clustering, both pixel color and position features are taken into account, and the logarithm mechanisim is introduced to balance the differences of their values. Through the special color distance filtering process the boundaries of superpixels are smoothed more effectively. Simulation results show that the LRRC algorithm is simple for use, efficient for computation, and can get a high boundary recall and a low under-segmentation error. When the number of partition superpixels is fairly large, the performance of the LRRC algorithm is better than other powerful superpixel algorithms in available.

关 键 词：K均值聚类 超像素 特征距离 对数机制 图像分割 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]