改进的半监督降维方法及其在人脸识别中的应用

Improvement semi-supervised dimensionality reduction method and its application in face recognition technology

在人脸识别算法中,维数较高的人脸图像数据容易引发“维数灾难”,增加数据的计算复杂性和冗余性,因此,提取低维人脸特征成为人脸识别方法的关键一步.为了在降维过程中保持数据之间的约束关系和内部结构信息,本文融合判别成分分析和主成分分析的思想,提出改进的半监督降维方法.通过改进优化函数的定义,在使用成对约束关系的传递性与排斥性的同时保持了数据的整体信息,将原问题等价转换为一个带参数的广义特征值问题.在数值实验中,利用ORL、YALE、BIO人脸数据库检验人脸图像的重构,比较几类数据降维算法的识别率,验证了改进的半监督降维方法的优越性.

In face recognition,image data of high dimension easily lead to"dimension disaster"and increase computational complexity and redundancy of data.To extracting low-dimensional face features is an extremely critical step in face recognition.Combining DCA and PCA with SSDR,an improved semi-supervised dimensionality reduction(SSDR)is presented,which uses both unlabeled data and prior information.The improved SSDR maintains the constraint relationship between data and the internal structure information of data in the process of dimension reduction.The definition of the optimization function is improved by using pairwise constraints between transfer and exclusion at the same time,and the overall information of data is kept.The problem of extracting face features is equivalently transformed to a generalized eigenvalue problem with parameters.The ORL,YALE and BIO face databases are used to test the face image reconstructions and compare the recognition rates of several dimension reduction algorithms.The experimental results indicate that improved SSDR method has certain superiority.