DOI:           10.5769/D2000001 and http://dx.doi.org/10.5769/D2000001

Author:       Paulo Quintiliano da Silva

Advisor:     Antonio Nuno de Castro Santa Rosa

University: University of Brasilia (UnB), Brazil

Year:          2000

To download this dissertation - Portuguese version, click here.

                                                      English version, click here. (soon...)

To download this dissertation presentation, click here.

To download this dissertation main Matlab programs, click here. (soon...)
To return, click here.

The recognition of well-known faces is very important in our social relationships. It is trivial function for our brain, however extremely important for our simpler and daily activities, because our relationship with other people is based on face recognition. Face Recognition can be divided in three different stages: face detection, facial representation and face recognition properly said. The proposed approach is based on PCA, decomposing the face images in a small characteristics group - the eigenfaces - based on Linear Algebra concepts - the eigenvectors and the eigenvalues. Thus, the eigenfaces are the principal components of the original face images, obtained by the PCA decomposition, forming the “Face Space” from these images. Face Recognition is obtained from two phases: training and recognition. The training stage is executed by means the projection of 4 face images of each class in the “Face Space”, generating a big dimension matrix Ok with as so much columns as the used classes. Recognition is executed by means the new face projection in the “Face Space”, generating the vector O. Thus, we calculate the euclidean distance between the vector O - that represents the new face - and each matrix Ok column - that represents the face classes. The found distance that is inside the threshold of a certain class and if it has the smallest value, it implies that there was the face recognition of the face belonging to the class of Ok. There are several problems that commit the Facial Recognition algorithms performance, as facial expressions, scale, inadequate illumination, disguises and faces position. The proposed approach is quite robust on facial expressions treatment and disguise with use of glasses. Some simmetryzation techniques are presented to improved a lot the performance of this algorithm based on eigenfaces, when working with images obtained in not controlled illumination conditions. he eigenface concept was expanded to the eigenfeatures: eigenmouth, eigennose and eigeneyes, in way to allow the construction of algorithms with a good effectiveness even working with halp-occluded images, facilitating the facial recognition just from face images fragments of approximately 20%. 


Face recognition, PCA, Principal Component Analysis, image processing, computer vision, pattern recognition, eigenfaces, eigenfeatures, eigenmouth, eigennose, eigeneye.