neural classifier ensemble using error-correcting output codes: access control application

abstract biometric access control is an automatic system that intelligently provides the access of special actions to predefined individuals. it may use one or more unique features of humans, like fingerprint, iris, gesture, 2d and 3d face images. 2d face image is one of the important features with useful and reliable information for recognition of individuals and systems based on this kind of feature, are fast, accurate and reliable. there is a vast literature for face recognition problem. one of the most interesting methods in pattern recognition recently gained high interest is combining classifiers. the main idea in this method is to use a number of simple classifiers instead of designing a huge classifier for solving a complex classification problem. it has two advantages: first point is that the task of designing simple classifiers is easy and has low computational cost. second, if a classifier makes a misclassification on special samples and others classify it correctly, the multiple classifier system (mcs) as a whole is able to make final correct decision. error correcting output codes (ecoc) is one of the most useful methods in building mcs. this method decomposes a multiclass problem into a number of simpler binary sub-problems called dichotomies. each dichotomy is tackled by a base binary classifier and subsequently using some existing reconstruction methods final class label is determined. in this thesis, we tackle automatic face recognition problem with the ecoc classifier. in this approach, a multi-class face recognition problem is decomposed to some simpler binary sub- problems. we use multilayer perceptron neural networks with error back-propagation learning algorithm as binary classifiers in ecoc. our proposed method, achieves some improvements in both decomposition and reconstruction stages. in the decomposition stage and for finding (semi) optimal code matrix for problem at hand, thinned ecoc coding method is proposed. for decision making in reconstruction stage, we propose ga-based and performance-based decoding. for the evaluation of different algorithms, we use orl face data set and face images taken in the unconstrained condition of our laboratory. these samples have variations in pose, expression and illumination. during the experiments, we investigate the effects of different factors such as number of training samples, binary classifier structure and illumination variation on the performance of the recognition system. experimental results show the robustness of the proposed recognition system in comparison with other existing methods.