GROUPED WINDOW-BASED NEURAL NETWORK APPROACH TO FACE RECOGNITION

Abstract

Face Recognition has been a topic, which a lot of scientists have tried to solve for many years. Automated recognition of human faces is an active research topic. Many systems based on different models, using different approaches have been developed for automated face recognition [1,2,3]. In this thesis, we are going to discuss an approach for face recognition using neural networks. We propose the Grouped Window-based Neural Networks (GWNN) [4] to solve this particular problem. A database of 100 face images was formed. These face images are taken by CCD camera with a resolution of 400 X 500, and 256 gray levels. These images were compressed to a resolution of 200 X 250 first, then they were Fourier analyzed to make them invariant with respect to translation, rotation and scaling. Each images was then cut into a few portions. Principal Component Analysis (PCA) was applied to each portion ( called a window) to obtain a set of parameters. These parameters were then fed into GWNN for analysis. In chapter one and chapter two, we will give an introduction of face recognition systems and describe other people's work. In chapter three, we present the mathematical description of Principal Component Analysis (PCA). In chapter four and chapter five, we describe the neural network mechanism, structure and its functioning. In chapter six, we discuss how to make face images become invariant with respect to translation, rotation and scaling by applying Fourier analysis. In the last two chapters, we present our experimental results. By combining Grouped Window-based Neural Networks technology and Principal Component Analysis, this system achieves an encouraging result of an average recognition rate of 88%, comparing to the recognition rate of 64% for non-window based system.