The Research Of Linear Subspace Method For Face Recognition

Linear subspace method is the mainstream technique for face recognition. It has been given more and more attention owing to its good properties such as strong describing ability, efficient computation and easy implementation. In this paper, we focus our attention on the theoretics and algorithms belonging to linear subspace method. After making a deep research , we develop some new algorithms as regards it and these algorithms are verified to be effective in the application of face recognition.We first explore the called small sample size (SSS) problem under the condition of high dimension samples, which is intractable and familiar when using Linear Discriminant Analysis(LDA) for face recognition. In this paper, three effective methods for linear analysis in singular case is researched, and they are the general framework for linear analysis algorithm, Maximum Margin Criterion(MMC) and Inverse Fisher Discriminant Analysis(IFDA). Given that IFDA is proposed without any theoretic base, we make clear the rationality of IFDA, and consequently theorise IFDA algorithm. The above three algorithms are verified in the application of face recognition using ORL and Yale face database, through which we discover parts of their characteristics.Another aspect attracts us about linear discriminant analysis is the computing efficiency. In this paper, our focus is on the image's compressing orientation. The face recognition experiments help us to dig out some significant and complicated attribute concerning the compressing scheme. Later we proposed a new distance metric for image feature and its validity is validated through experiments as well.Independent Component Analysis(ICA) is another important subspace method for face recognition besides LDA. In this paper, we discover the speciality of ICA method when used for small class problem. Then we prove the conclusion that ICA using PCA as the baseline algorithm on a new way , and explore the performance of ICA under the condition of feature selection. Another significant work about ICA in this paper is that we develop a novel ICA algorithm for face recognition which can utilize the class information of training samples, and it is verified to be effective in the application of face recognition.At last, we proposes a novel algorithm called Supervised Kenel PCA(SKPCA) as a supplement to subspace analysis for face recognition. SKPCA combine supervise learning ideology with KPCA algorithm, so result in relatively strong classification capacity. We verify SKPCA in application of face recognition on ORL and Yale face database and find that SKPCA can reach better recognition performance comparing with KPCA and PCA.