Face recognition using Orthogonal Transforms and Vector Quantization Techniques

Abstract

This thesis is an attempt to unravel the problem of human face recognition. Face recognition is a biometric authentication method that has become more and more relevant in the recent years. Face recognition is a popular research area where there are different approaches studied in the literature.In this thesis face recognition problem is handled by applying Principal Component Analysis (PCA), Various Orthogonal transforms and different Vector Quantization (VQ) codebook generation techniques. The eigenvectors are the principal component of the set of images. Each face can be closely represented by a linear combination of the Eigenfaces. The Eigen face method tries to find a lower dimensional space for the representation of the face images. Here to reduce the dimensionality instead of utilizing all the Eigenfaces for recognition purpose only small number of Eigenfaces is selected according to the energy contain which is close to total energy of original face. This particular method helps to reduce the dimensionality efficiently. The main drawback of PCA is scalability. As dataset changes the whole eigenspace distribution also changes. To avoid this difficulty various orthogonal transforms like Discrete Cosine Transform, Discrete Sine Transform, Walsh Hadamard Transform, Slant Transform, Wavelet Transform and newly proposed Kekre s Transform are applied on controlled standard ORL database and unconstrained local database prepared by using Indian origin faces of 50 male and 50 female. The concept of image energy compaction in low frequency coefficients is explored here. Instead of using total image energy and all the transformed coefficients for comparing a test image with the stored database only selected number of coefficients according to desired energy of image is utilized for recognition purpose. To reduce computational time and complexity the separablity property of transforms is utilized.