LDA/GE在人脸识别中的应用  

LDA/GE for face recognition

作　　者：李子荣[1] 杜明辉[1] 

机构地区：[1]华南理工大学电子与信息学院,广州510641

出　　处：《计算机工程与应用》2008年第25期179-181,185,共4页Computer Engineering and Applications

基　　金：广东省自然科学基金No.05006593~~

摘　　要：基于谱图理论和流形学习,同时受FKT的启发,LDA算法可以进一步改进和化简。FKT已经被证明为二次判别分析中的低秩近似最优解,一开始只是用于二类识别问题,近年陆续有文章将它用于人脸识别中,以解决样本数小于样本维数的问题。LDA使用FKT求解的时候,在图嵌入的框架下可以转化为两次嵌入求解,第一次嵌入是PCA,第二次是由PCA的特征向量张成的空间中求判别式中分子的特征值分解问题。这样不仅去除了小样本问题下的奇异性困扰,更重要的是,基于谱图理论,将判别分析中的除法去掉了。最后,给出和其他方法比较的人脸识别实验结果和结论。Based on spectral graph theory and manifold learning,and inspired by Fukunaga-Koontz Transform,the traditional LDA is simplified and improved.FKT has been proved to be the best low-rank approximation to Quadratic Discriminant Analysis.The transform is only used in the two-class classification problem at the early time,and recently has been used in face recognition to solve the Small Sample Size Problem.LDA can transform to a two-stage graph embedding,first it is the PCA,then eigenvalue decomposition of the numerator of the Discriminant in the spaces spanned by the principle eigenvectors of PCA.Both the singularity of the data and the ratio form in Discriminant Analysis are removed.

关 键 词：降维 流形学习 图嵌入 人脸识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]