基于高维特征和稀疏子空间聚类的图像分割方法研究

项目名称： 基于高维特征和稀疏子空间聚类的图像分割方法研究

项目编号： No.61472303

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 王卫卫

作者单位： 西安电子科技大学

项目金额： 80万元

中文摘要： 图像分割是图像理解与识别的基础，是计算机视觉亟需解决的问题。传统图像分割方法使用个别图像特征，个别特征不能反映自然界的丰富多样性，分割精度有限。本项目将图像分割看成图像高维特征的聚类问题，并利用目前高维数据聚类中最有效的、无监督、无需先验信息的稀疏子空间聚类方法来解决。创新点有: 以子空间表示矩阵应具有类间稀疏类内一致特性为原则，利用分组p-q范数和迭代重加权稀疏范数设计正则项，将核范数推广为奇异值的非凸p范数，将其与新的稀疏度量相结合，得到类间稀疏类内均匀的子空间表示；在分析新正则项的数学性质基础上，给出类间稀疏类内一致正则项的条件，为设计类间稀疏类内一致正则项提供理论依据，在此基础上设计新的正则项；分析特征域噪声特性，设计与之相适应的误差度量和数据项；利用推广的主元素分析法和本项目设计的正则项、数据项建立图像高维特征降维和子空间表示相联合的模型与算法。预期取得与国际先进水平相当的成果。

中文关键词： 图像分割；高维特征；稀疏子空间聚类；谱聚类

英文摘要： Image segmentation is a fundamental problem in image interpretation and recognition, which are important parts in computer vision research. For simplicity in modeling image segmentation, most traditional methods just use a few features, which, however, cannot well characterize the variety of the nature, thus having limited segmentation precision. In this proposal, we take image segmentation as a clustering problem of high dimensional features，and use the idea of sparse subspace clustering to solve it because the sparse subspace clustering is one of the most effective methods for segmenting high dimensional data. The contribution of this proposal includes: we propose the priciple that the subspace representation matrix should be sparse between clusters and uniform within clusters.In order to enforce the inter-cluster sparsity and intra-cluster uniformity of the subspace representation matrix, we propose to use the so called p-q norm and iterative reweighted 1-norm or 2-norm to design regularization terms. we also generalize the nuclear norm to the p-norm of the singular values, and combine it with the new sparse measures to design regularity terms. We will explore the mathematical properties of these regularization terms and present the general conditions on the regularization term to enforce the inter-cluster sparsity and intra-cluster uniformity of the subspace representation matrix. The conditions will provide priciples on designing new regularization terms. We will design new error measurement and data terms according to the statistics of noise in image features. Finally, by using the extended PCA methods, the regularization terms and data terms we design to incorporate the feature dimension reduction into the subspace representation model such that the computation load be reduced and the method be more robust. We expcet to achieve results of comparative quality with internatioanl level.

英文关键词： image segmentation;high dimensional feature;sparse subspace clustering;spectral clustering