AbstractFace recognition is an important subject in authentication systems and
computer vision. In this thesis two methods have been proposed for face
recognition.
The first method presents a feature extraction method based on Zernike
moments. Five power transform distance metrics are constructed for the
purpose of defining similarity measures for holistic face recognition. The
power transform metrics are based on the logarithm of five distance metrics:
Euclidean, City-Block, Soergel, Lorentzian and Minkowski metrics. The
constructed power logarithmic metrics have been called PL-Euclidean, PLCity-Block, PL-Soergel, PL-Lorentzian and PL-Minkowski. Distance is
considered in the Zernike domain of face images, a domain that contains a
vector of Zernike moments for each face image.
The second method represents a new approach for face recognition that
incorporates Prewitt edge detection, Gabor filter and Zernike moments. The
five distance metrics of the Schoenberg transform are constructed for the
purpose of defining similarity measurements for holistic face recognition in
the Zernike domain. The Schoenberg transform metrics used the logarithm of
distance metrics: Minkowski, City-Block, Euclidean, Soergel, and Lorentzian
metrics. The constructed Schoenberg logarithmic metrics called SLMinkowski, SL-City-Block, SL-Euclidean, SL-Soergel, and SL-Lorentzian.
These distance metrics are further modified as similarity measures then
normalized to be restricted over the range (0,1) so that a fair comparison
could be made with the well-known image structural similarity measure
(SSIM) and Feature Similarity Index for Image Quality Assignment (FSIM).
The purpose is discovering the similarity of a given face image with a facedatabase to check for best match.
A measure for quality of recognition is also proposed. This confidence
measure is based on the difference between similarity given by the most likely
match with that given by the second-likely confusing face.
Simulation results show that the proposed distance measures have
superior performance as compared to classical metrics, and the proposed
similarity measures are all superior to the well-known SSIM and FSIM.