Performance Enhancement of Subspace Learning Face Recognition by Effective Use of Classifiers

Abstract

Subspace based algorithms belong to one of the most explored face recognition algorithm categories which follow a holistic approach for feature extraction. These methods operate directly on the pixel intensities of a facial image and extract features. The basic trait of these algorithms is that they reduce dimensionality to reduce the computational complexity of feature extraction while keeping the statistical separation between different classes. Therefore these algorithms are the economical choice for feature extraction. These algorithms are based on the key concept that most of the information in a facial image is highly redundant and that the discriminating features reside in a subspace of the face image. Therefore these algorithms aim to extract these features by reducing the redundant and non-discriminating information. The choice of a classifier is the key factor in designing an efficient pattern classification system. This choice very closely relates to the data on which it is going to be applied. Another important issue is the irrelevancy in reported results of different classifiers. The evaluation criterion which is set for evaluating a specific classifier plays a significant role in determining the true potential of a proposed classifier. There is a need to evaluate these reported classifiers using the same evaluation criterion to judge the suitability of each classifier for a specific imaging condition. For face recognition, a surfeit of classifiers has been proposed to date but none of them alone is capable enough to cater with all the inherent variations of the facial image data. Therefore there is a need to explore combinations of classifiers known as ensemble classifiers. As different classifiers extract complementary features of the object to be classified, therefore combining the properties of individual classifiers in an ensemble classifier does result in increased classification accuracy. The overall suitability of this ensemble classifier depends on the memory and computational complexities of the constituent base classifiers. VI In this thesis, a newly reported and highly cited face recognition algorithm Laplacianfaces is initially explored for its true potential by varying its internal and external parameters for different face recognition tasks. Based on the outcome of this initial analysis, other famous subspace face recognition algorithms are also evaluated by using distance metrics both from the image space and mahalanobis space. This evaluation was performed by using the evaluation methodology employed in Face Recognition Vendor Tests (FRVT) and FERET evaluations. These algorithms are evaluated against various probe sets from three different and famous facial databases namely FERET, ORL and YALE. This study hence provides enough testing variables to judge the performance of algorithms against different imaging conditions or facial variations. Based on this exhaustive comparative analysis, a group of six most accurate and most economical classifiers are selected. Ensemble classifiers with combinations ranging from two to six of these best selected base classifiers are evaluated against the same testing conditions. The ensemble classifiers are constructed by combining base classifiers using two simple ensembling techniques namely re-ranking and weighted scoring approach. The average performance of this ensemble classifier also called unified classifier is found to be well ahead of that for the individual constituent base classifiers. The work reported in this study proves the effectiveness of ensemble classifiers for face recognition tasks. The results of the proposed unified classifier in comparison to the best performing subspace algorithms demonstrate that the unified classifier has a global performance and can handle different variations effectively.