摘要
矩阵分解算法是模式识别中一种常用的图像表示方法.针对传统的矩阵分解算法不能提取数据本质结构的问题,提出一种局部敏感的稀疏概念编码的图像表示算法.在基向量学习时,利用局部敏感鉴别分析方法提取样本的几何结构和判别信息,使得学习到的基更能体现数据的高层语义结构信息;然后对每个样本在基向量上进行稀疏表示学习,得到样本的表示系数;最后对样本进行表示与分类.在COIL20和ORL数据库中的实验结果表明,与其他几种矩阵分解算法相比,文中算法聚类的准确率和互信息得到了有效的提高,验证了其有效性.
Matrix Factorization is very effective image representation approach in pattern recognition. The traditional matrix factorization methods cannot capture the intrinsic structure information. In this paper, a novel method, called Locality Sensitive Spare Concept Coding （LSSCC）, is proposed which can capture the intrinsic geometrical structure and discriminate information in basis learning. Therefore, it can find a basis set capturing high-level semantics information of the data. And the coefficients are obtained when the samples are sparse representation on the basis vectors. Finally, the samples are represented and classified. The clustering experiments on the COIL20 and ORL database demonstrate the proposed algorithm can effectively improve the accuracy and normalized mutual information in clustering and verify the effectiveness compared to other matrix factorization algorithm.
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2014年第1期81-87,共7页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金(61272220
61101197)
国家自然科学基金重大研究计划重点支持项目(90820306)
关键词
矩阵分解
局部敏感
稀疏概念编码
几何结构
判别信息
matrix factorization
locality sensitive
sparse concept coding
geometrical structure
discriminate information
