The human face is one of patterns to identify a human beings and the automatic face recognition system has attracted considerable attention in recent years. In this thesis, two new face recognition paradigms are studied. One is based on Kohonen``s Self-organizing Maps(SOM) and the other is based on isodensity maps.
In the SOM method the facial components are extracted using model-based vision and the feature vectors of each component are derived from frequency domain without assuming any feature points of the facial components. Therefore, difficulties in extracting the exact feature points are removed in this paradigm. Then, the SOM of four facial components such as mouth, right eye, left eye, and nose are applied to classify the unknown face. In the learning phase, each SOM is tuned by learning vector quantization to minimize the classification error. In the recognition phase, the decision processes are performed for each SOM independently, and the obtained results are totally utilized to identify unknown face.
In the method using isodensity maps, the isodensity maps are obtained by the histogram of facial image and their characteristics are extracted by geometrical feature analysis. As a representation method of extracted attributes, topological graphs are introduced and a simple deterministic graph matching is used to recognize unknown faces from stored database.
At last, the effectiveness of two proposed methods are verified by experimental results and some typical applications are also introduced after comparing the two methods and analyzing each face recognition paradigm respectively.