Face Recognition Enhancement for Security Purposes

Abstract

With the fast evolution in security fields in ministries and large firms, face recognition systems are needed to identify and recognize people. Given an input face image and a database of face images of known individuals, face recognition systems aims to verify or determine the identity of the person that in the input image. Face Recognition is a task that human vision system seems to perform almost effortlessly. Yet the goal of building computer-based systems with comparable capabilities has proven to be difficult. Many techniques for face recognition have been proposed over the last 30 years, and some were even proposed earlier. Some techniques extract local features from the face image (structural matching techniques), while others uses the face image as a whole (holistic matching techniques). Some techniques are a hybrid of both. Subspace techniques which belong to the holistic category are the most successful techniques for face recognition. These techniques aim to reduce the dimensions of the face images by projecting the high-dimensional images to a lower-dimensional subspace, where each image is represented by a linear combination of vectors in the subspace. In this work, a fully face recognition system to classify faces has been implemented. Principal Component Analysis (PCA) technique has been used as one of the holistic matching techniques for dimensionality reduction and feature extraction. Using PCA, features are extracted from the whole face, then an empirical method to choose the most principal “important” features from the face has been developed, this empirical method helps the classifier to recognize faces more easily and more efficiently. The mathematical model of PCA technique is proposed in this thesis, and its equations has been performed and implemented using Matlab Programming. After the feature extraction phase, a Feed Forward Neural Network is used for classification, specifically, a Multi Linear Perceptron (MLP) networks is used, and a Back Propagation (BP) algorithm is used as a learning algorithm. An empirical method has been developed to choose the number of hidden neurons of the hidden layer to avoid both underfitting and overfitting problems. The mathematical model of the back propagation algorithm is proposed in this thesis, and has been performed using the neural network tool box available in Matlab. The Olivetti Research Laboratory (ORL) database is used for performance evaluation.