|
Abstract:
|
This thesis concerns the field of facial feature recognition and evolutionary computation. The work presented here is part of the work that was carried out in the EU MUHCI (MultimodalHuman Computer Interaction) project. We present a novel approach to the problem of facial feature recognition, building a learnable multi-agent system which is trained with the use of evolutionary computation. The aim of this thesis was to design and implement a general multi-agent system in C++ able to solve various problems through the learning of the agents comprising the system. The learning rules used are based on the paradigm of evolutionary computation which in turn is based on the creation of a population of agents using the rule of natural evolution, i.e. "survival of the fittest".The idea of using an approach based on multi-agent systems and evolutionary computation for facial feature recognition vs. the traditional approaches is due to the variabihty of visual facial images. To build a model for the class which we recognize as a "face" which would be ap plicable to every person, pose and illumination condition among others, seems to be very difficult the same occurs with the classes "eye", "nose", "lip", etc. However, humans do it everyday. Phe way humans classify is based on to some degree learning and experience. The system constructed here is applied to the face recognition problem, using as inputs Gabor features of the images, which are well known as a good characterization of the features in the face and an annotated visual face database. The agents perform simple operations based on the integrate-and-fire operation of the artificial neuron. Each agent is rewarded according to how close its outp ut is to the desired output. Finally, the selection in the population is done based on the value of the agent estimated in this way. After enough traimng the agents in the population should be able to recognize the feature that they were trained to. The tests done in this thesis indicate the practicability of using evolutionary computation for the recognition of facial features. /Kir10 |