Kernel variants of extended locality preserving projection

Abstract

In recent years, non-linear dimensionality reduction methods are getting popular for the handling image data due to non-linearity present in data. For the image recognition task, non-linear dimensionality reduction methods are not useful as it is unable to find the out-of-sample data representation in the reduced subspace. To handle non-linearity of the data, the kernel method is used, which find the feature space from higher dimensional space. One can find the reduce subspace representation by applying the linear dimensionality reduction techniques in the feature space. Extended Locality Preserving Projection(ELPP) tries to capture non-linearity by maintaining neighborhood information in the reduce subspace but fails to capture complex-nonlinear changes. So kernel variants of ELPP are proposed to handle non-linearity present in the data. This thesis addressed kernel variants of the ELPP which efficiently handle the complex non-linear changes of the facial expression recognition. The proposed kernel variants of the ELPP is applied for face recognition on some benchmark databases. Proposed approaches are also able to remove the outlier present in the data.