非一致相似测度的图表示多观测样本分类算法  

Multiple Observation Sets Classification Algorithm Based on Graphical Presentation of Inconsistent Similarity Measure

作　　者：胡正平[1,2] 赵艳霜[1] 荆楠[1,2] 

机构地区：[1]燕山大学信息科学与工程学院,秦皇岛066004 [2]河北省计算机虚拟技术与系统集成重点实验室

出　　处：《信号处理》2012年第11期1587-1594,共8页Journal of Signal Processing

基　　金：国家自然科学基金(61071199);河北省自然科学基金(F2010001297);第二批中国博士后科学基金特别资助(200902356)

摘　　要：多观测样本分类问题中,样本表示成流形上的点,针对如何利用多观测样本的流形结构提高其分类性能的问题,提出非一致相似测度的Graph表示多观测样本分类算法。首先综合数据的全局与局部结构特性,构造一个非一致相似测度,非一致相似测度主要考虑类内和类间差别,能有效地体现数据实际聚类的分布特性;其次构造非一致相似测度Graph,进而得到样本之间的相似度矩阵,然后通过一个格拉斯曼联合核把最佳投影的计算转化成寻找瑞利熵的最大特征向量问题,进而得到投影矩阵。最后将本征流形上的点投影到另一个流形上,使用最近邻分类器完成分类。在ETH-80物体识别数据库、CMU-PIE人脸数据库和BANCA数据库上进行对比实验,实验结果表明该方法优于传统方法。In classification problem of multiple observation sets, the samples are represented as points on Grassmannian manifolds, with regard to how to exploit the manifold structure to improve the classification performance, a multiple observation sets classification algorithm based on graphical presentation of inconsistent similarity measure graph is presented. First of all, considering the characters of global and local data structure comprehensively, an inconsistent similarity measure is constructed, which regards the distinction of within-class and between-class mainly and can effectively reflect the distribution character of actual data clustering. The second step is to obtain the similarity matrix via inconsistent similarity measure graph, after that, the computation of the optimal map is transformed into the search problem of the largest eigenvectors of the Rayleigh quotient by a combined Grassmannian kernel and then the projection matrix is obtained. Lastly, points on the manifold can be mapped into another space, the final classification is completed exploits the nearest neighbor classifier. Three comparative experiments are conducted on ETH-80 object recognition dataset, CMU-PIE and BANCA face recognition datasets, the results prove that the algorithm performs better than traditional algorithm.

关 键 词：模式识别 非一致相似测度 图表示 格拉斯曼流形 最近邻分类器 多观测样本分类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]