In partial fulfillment of the Requirements for the Degree of
Master of Science
Joseph Liang
will defend his thesis
Eye Detection and Face Recognition:
Comparison of Algorithms
Abstract
Face recognition has a wide range of applications such as personal
identification and authentication, criminal identification,
security and surveillance, image and film processing, and
human-computer interaction. Although many methods exist, this
thesis compares two recent face recognition algorithms - the
Regularized Direct Quadratic Discriminant Analysis (RD-QDA) method
proposed by Lu et al. and the Direct Linear
Discriminant Analysis (D-LDA) method developed by Yu and Yang.
In terms of accuracy, the D-LDA algorithm could not
match the performance of the RD-QDA algorithm because the RD-QDA
algorithm is an optimized discriminant analysis method. However,
in order to find the optimum regularization parameters of RD-QDA,
an exhaustive search must be performed, which is very
computationally demanding. Thus, D-LDA is quicker, but RD-QDA is
more accurate.
Since the first step of face recognition is to locate the face,
this thesis also compares two eye localization algorithms that
will be used to help locate the faces in the images. The two eye
localization algorithms are the Efficient Face Candidates Selector
(EFCS) method by Wu and Zhou and the Generalized
Projection Function (GPF) method proposed by Zhou and Geng.
In summary, the EFCS algorithm is good at
approximating the eye locations, and the GPF algorithm is more adept at
taking that approximation one step further in obtaining the precise
eye locations.