Non - Linear Feature Extraction And Face Recognition Based On Kernel Method

Feature extraction is a key issue of face recognition, and the effective feature information has a remarkable influence on recognition performance. This paper makes focuses on kernel-based nonlinear feature extraction methods and propose some novel algorithms for feature extraction. All the algorithms are verified to be effective in the application of face recognition. This paper mainly discussed the following questions:Kernel Principal Component Analysis(KPCA) is an effective nonlinear feature extraction algorithm as an extension for classical PCA method. However, KPCA has its limitation by reason of ignoring the class label information of samples. In this paper, Class-information-incorporated Kernel PCA method is analyzed deeply, then an amelioration of nearest-neighbor classifier is made based on the analysis. Finally, the experimental results on Iris dataset and several face databases indicate that the classifier is more effective for classification and validate the theoretical analysis.Kernel Fisher Discriminant Analysis(KFDA) and Kernel Maximum Margin Criterion (KMMC) are effective feature extraction methods, which have been widely studied and is suitable for face recognition. This paper makes an in-depth analysis of the two feature extraction methods in theory. Then combining with local geometric information of samples belonging to the same class, a novel algorithm named Local Kernel Maximum Margin Criterion(Local KMMC) is proposed. Local KMMC uses both the local distance information and class information of training samples, so can extract non-linear features more adapted to classification.Kernel Canonical correlation analysis (KCCA) is a powerful tool to analyze the underlying dependency between two data sets. Kernel Generalized Canonical Correlation Analysis(KGCCA) and Kernel Discriminative Canonical Correlation Analysis(KDCCA) improve KCCA on one aspect. In this page, Kernel Generalized Discriminative Canonical Correlation Analysis is proposed, Which embodies the impacts of both within-class correlation and between-class correlation on classification. The experiments show the superiority of KGDCCA to other relative methods in terms of the recognition performance.In the end, an algorithm called Local Discriminative Canonical Correlation Analysis(Local DCCA) is proposed. By means of constructing distance matrix of samples belong to the same class, Local DCCA aims to maximize the intra-class correlation metric meanwhile minimize the inter-class correlation metric. Using the kernel trick, nonlinear extension of Local DCCA(Local KDCCA) is raised. The experimental results tested on several face databases indicate that these proposed methods are effective for face recognition task.