摘要
提出一种基于广义的2DLDA算法,简称:G2DLDA.首先,由于2DLDA算法提取的特征向量矩阵S-1wSb通常不是标准正交特征向量矩阵,因此该方法会严重影响特征提取的质量.本文根据Sw矩阵是对称正定的,即:具有Sw=S1/2w×S1/2w性质,将2DLDA算法的特征向量矩阵转化成基于标准正交特征向量矩阵,即:S-1/2wSbS-1/2w.其次,G2DLDA算法与2DLDA一样不会产生小样本事件,因为方程式S-1/2wSbS-1/2wv=λv的右端为单位矩阵,是满秩的.最后,G2DLDA算法采用基于Cosine-范数度量方式进行分类,实验证明该度量方式优于其他度量方式,如:欧氏距离度量方式以及F-范数度量方式.在实验阶段,本文采用Yale、ORL和JAFFE三个数据库对该算法进行测试与分析,实验结果证明该算法具有较好的鲁棒性,同时能够获得较高的识别率.
In this paper, a novel method based on the generalized 2DLDA, G2DLDA for short, is proposed. First of all, it will affect the quality of feature extraction due to the eigenvectors Matrix S-1w Sb of 2DLDA method being usually not orthogonal. However, Sw matrix is symmetric and positive definite, namely Sw = S1/2w×S1/2w, so we can change the eigenvectors Matrix to standardized orthogonal vectors Matrix,namely S-1/2wSbS-1/2w. In the squeal,because the right side of equation S-1/2wSbS-1/2wv = λv is unit matrix which is full rank, G2DLDA as well as 2DLDA will never produce the ＂Small Simple Size＂. In last but not least,G2DLDA method using Cosine-norm metric for classification is better than other metrics such as the Euclidean distance metric and the F-norm metric through experimental analysis. In experimental phase, Yale Face Database,ORL and JAFFE Database are used by the test and analysis. Experimental results demonstrate that our method has better robustness and higher recognition rate than the state-of-art methods.
出处
《小型微型计算机系统》
CSCD
北大核心
2015年第4期856-861,共6页
Journal of Chinese Computer Systems
基金
国家自然科学基金重点项目(60736008)资助
国家自然基金青年科学基金项目(60903141)资助
北京市自然科学基金项目(4122017)资助
北京教委暨北京自然基金重点项目B类(KZ201210028036)资助
