FACE RECOGNITION USING NEURAL NETWORKS

Abstract

Face recognition is a computer application of recognizing persons by their facial images, which may vary with makeover, expressions, age or due to noise added during image acquisition or during transmission over a network. Work on Face recognition is going on from around last twenty years, and has got immense future scope and great commercial value. Face recognition is being used for Access Control, Protection and Security, Industrial Automation and Robotics etc. Face recognition system is based on principals of various fields like Image Processing, Linear Programming, Statistics and Neural Networks etc. Various Image Processing Techniques like Laplacian, Laplacian of Gaussian, and Histogram Equalization etc. are used for image preprocessing stage to increase sharpness and dynamic range. Facial images contain redundancies which do not help in recognition but increase processing overhead, Techniques like Fourier Transform and Wavelet Packet Decomposition are used for feature generation or dimensionality reduction. _Concepts from Linear Programming and statistics like Principal Component Analysis, Singular Value Decomposition, and Independent Component Analysis etc. are also used for generating features of reduced dimensions from original data, which has same recognition potential as original data. Neural Networks are powerful tool for recognition type of computation. They are trained on representative set of images, once trained they can recognize images they have never seen. Most important feature of such networks are that they tolerate noise extremely well. In this dissertation an algorithm for improved Face Recognition has been developed. Laplacian has been used for preprocessing the facial images. Singular Value Decomposition has been used for feature generation dimensionality reduction. Neural Networks have been used ' for recognition purposes. In order to improve noise tolerance of developed face recognition system Neural Networks has been trained by deliberately generated noisy data. Further, in training Neural Networks a technique Resilient Backpropagation has been used which made the training faster. Face Recognition system developed has an accuracy of 91% under zero random noise condition. Further a study of variation of accuracy with noise has been carried out and results are presented in chapter 6.