Abstract

ABSTRACT
Two-Dimensional Face Recognition Algorithms
in the Frequency Domain

Alper Serhat Zeytunlu

The importance of security, law-enforcement and identity verification has necessitated the development of automated stable, fast and highly accurate algorithms for human recognition. Face recognition is one of the most popular techniques used for these purposes. Face recognition algorithms are performed on very large size of datasets obtained under various challenging conditions. Principal component analysis (PCA) is a widely used technique for face recognition. However, it has major drawbacks of (i) losing the image details due to the transformation of two-dimensional face images into one-dimensional vectors, (ii) having a large time complexity due to the use of a large size covariance matrix and (iii) suffering from the adverse effect of intra-class pose variations resulting in reduced recognition accuracy. To overcome the problem of intra-class pose variations, Fourier magnitudes have been used for feature extraction in the PCA algorithm giving rise to the so called FM-PCA algorithm. However, the time complexity of this algorithm is even higher. On the other hand, to address the other two drawbacks of the PCA algorithm, two-dimensional PCA (2DPCA) algorithms have been proposed.
This thesis is concerned with developing 2DPCA algorithms that incorporate the advantages of FM-PCA in improving the accuracy and that of 2DPCA algorithms in improving the accuracy as well as the time complexity. Towards this goal, 2DPCA algorithms, referred to as the FM-r2DPCA and FM-(2D)2PCA algorithms, that use Fourier-magnitudes rather than the raw pixel values, are first developed. Extensive simulations are conducted to demonstrate the effectiveness of using the Fourier-magnitudes in providing higher recognition accuracy over their spatial domain counterparts. Next, by taking advantage of the energy compaction property of the Fourier-magnitudes, the proposed algorithms are further developed to significantly reduce their computational complexities with little loss in the recognition accuracy. Simulation results are provided to validate this claim.