Use of Experiment Design Methods to Determine Unseen Data in Face Recognition Problems.

Abstract

Face detection is the first step in any face recognition system. The purpose is to
localize and extract the face region from the background that will be fed into the face
recognition system for identification. This project will use general preprocessing
approach for normalizing the image and Radial Basis Function (RBF) Neural Networks
will be used for distinguishing face and non-face images. Face and non-face data will be
used for training the RBF network in order for the network to discriminate face and
non-face images. The non-face data were normally taken randomly from the internet or
subtracted from scenery images. Creating these non-face images is tedious especially
when thousands of data needed. Experiment design approach are investigated to solve
this problem where the non-face images are computer generated. However, this
approach was found to be unfeasible due to the long computational time to produce
one single non-face image even though a high performance computer was used. The
second focus of this project is to design a novel RBF neural network algorithm that can
detect non-face images effectively from a given sample image. In this project, an RBF
neural network using 200 number of centres and using a gaussian spread value of 5 gave
the best result in terms of face detection rate, discriminative result, small FAR and FRR
as well as the system can detect all faces in a test image commonly used in this research
area without indicating a false negative.