基于特征分组与特征值最优化的距离度量学习方法  

Distance Metric Learning Based on Feature Grouping and Eigenvalue Optimization

作　　者：赵永威[1] 张蕾[2] 李弼程[1] 王挺进[1] 吕清秀[1] 

机构地区：[1]解放军信息工程大学信息系统工程学院,郑州450002 [2]郑州升达经贸管理学院,郑州451191

出　　处：《数据采集与处理》2015年第4期830-838,共9页Journal of Data Acquisition and Processing

基　　金：国家自然科学基金(60872142;61301232)资助项目;全军军事学研究生课题资助项目

摘　　要：主流的距离度量学习方法都需要求解半正定规划(Semi-definite programming,SDP)问题,而其中每次循环迭代中的矩阵完全特征分解运算使得现有方法计算复杂度很高,实用性不强,难以应用在大规模数据环境。本文提出了一种基于特征分组与特征值最优化的距离度量学习方法。引入特征分组算法,根据特征各维数之间相关性对图像底层特征进行分组。在一定的约束条件下,将求解SDP问题转化为特征值最优化问题,在每次循环迭代中只需计算矩阵最大特征值对应的特征向量。实验结果表明该方法能有效地降低计算复杂度,减少度量矩阵的学习时间,并且能取得较好的分类结果。The current mainstream distance metric learning approaches that all need to solve the positive semi-definite programming problem（SDP）will lead to high computational complexity,and they are thus difficult to be applied to large scale datasets well because of fully matrix characteristics decomposition operational in each loop iteration.A distance metric learning method based on feature grouping and eigenvalue optimization is proposed considering the above problems.Firstly,a feature grouping algorithm is introduced to segment image features into several groups according to the correlations between each dimension of characteristics.Then,the SDP problem can be covered to eigenvalue optimization issue under some certain constraints.Therefore,only the maximum eigenvalues of matrix is needed in every loop iteration.Experiment results indicate that the computational complexity and the learning time of metric matrix are reduced effectively.Besides,the classification results are improved compared with the traditional methods.

关 键 词：距离度量学习 半正定规划 特征分组 特征值最优化 度量矩阵 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]