判别割(Dcut)的图像分割及其快速分割算法  

Fast image segmentations of Dcut

作　　者：邹小林[1,2] 陈伟福[1] 冯国灿[1] 

机构地区：[1]中山大学数学与计算科学学院,广州510275 [2]肇庆学院数学与信息科学学院,肇庆526061

出　　处：《中国图象图形学报》2012年第2期222-228,共7页Journal of Image and Graphics

基　　金：国家自然科学基金项目(60975083;U0835005)

摘　　要：谱聚类算法在模式识别和图像分割中得到了广泛应用。谱聚类算法能在任意形状的样本空间上聚类且收敛于全局最优解。采用一个新的谱聚类算法Dcut进行图像分割。Dcut完全满足聚类算法的一般准则:类内样本间的相似度大,类间样本的相似度小,因此Dcut在图像分割方面比Ncut具有更好的分组性能。为了克服Dcut分割速度慢,提出基于子空间的Dcut(SDcut)和基于分块的SDcut(BSDcut)两种快速算法。SDcut和BSDcut这两种快速算法具有Dcut的分组性能的同时,降低了分割图像的计算复杂度。通过对纹理图像和真实图像的分割,验证了新算法的有效性。Spectral clustering algorithms have wide applications in pattern recognition and image segmentation. They can cluster samples in any form of the feature space and have global optimal solutions. In this paper, a new graph-based spectral cluster algorithm called Deut is applied to image segmentation. Dcut completely satisfies the general criterion of the cluster algorithms：maximizing the within-cluster similarities while minimizing between-cluster associations. Compared with Ncut, Dcut has better grouping performance in image segmentation. In order to overcome Dcut＇ s shortcoming i. e. slow speed for image segmentation,two fast Dcut algorithms, i. e. subspaee-based Dcut （SDcut） and block-based SDcut （ BSDeut）, are proposed. SDcut and BSDcut have Dcut＇ s grouping performance whihe at the same time reducing the computational complexity.Experiments based on texture images and real images demonstrate the advantages of the proposed algorithms.

关 键 词：谱聚类 Dcut SDcut BSDcut 子空间 图像分割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]