Abstract:
The face detection problem is to identify the presence of a human face in an image. The problem has important applications to automated security systems, lip readers, indexing and retrieval of video images, videoconferencing with improved visual sensation, and artificial intelligence.
In this dissertation, Color based Neural Network and Appearance based techniques are examined and implemented for detecting frontal view human faces in color and gray scale images respectively.
In Neural Network based systems for detecting human faces in color images two approaches are used that vary in type of input i.e. RGB and YES color space; fed into the network. It is a color-based technique combined with unsupervised learning, or training, used to set the weights of the network. The idea for the network is to learn a chroma chart from a training set. Each system is trained on the same image set using the Levenberg-Marquad method. Same training images and test images are used to compare the results obtained from two different color spaces as input.
In appearance-based technique, face detection problem is divided into two parts; the first one feature extraction and other as classification. For feature extraction, EM algorithm of PCA (principle component analysis) is used for training and K-NN (K nearest neighbor) method is used for classification purpose where K is typically taken as 1. This is the advanced pattern classification based technique but it is typically used only for gray scale images with better performance.
These two techniques are implemented in MATLAB Version 6.5 and results are presented in chapter 4. In NN based technique 20 training images were taken and they take 1 minute for training. For classification based technique 4000 training images for face and 4000 nonface training images were taken and they take only 2-3 seconds. Comparison of these two .techniques was carried out on the basis of complexity, computational cost, training time, application areas, dependence on color space etc. The study finally reveals that appearance based technique is superior for most of the applications.
111
