关键词: 谱聚类, 图像分割, 相似度度量, 超像素, 协方差, NJW算法

Abstract: Considering the low accuracy of image segmentation method of traditional spectral clustering, an improved similarity measure of spectral clustering is proposed. Firstly, an image is made up of some superpixels by the pre-process of superpixels segmentation algorithm, and a graph based on superpixels is constructed. Secondly, similarity matrix is obtained by the similarity calculation of superpixels, which fully considers the features of superpixels including covariance descriptor, color information, texture information and edge information. Finally, NJW algorithm is used to segment the graph based of superpixels. Compared with current unsupervised segmentation algorithm, a lot of experiment results show that the proposed approach has higher segmentation accuracy. Besides, the object marked by user can be segmented precisely using proposed approach.

Key words: spectral clustering, image segmentation, similarity measure, superpixels, covariance, NJW algorithm