Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

Combining Principal Component Analysis and Neural Networks for the Recognition of Human Faces. A Case Study for Man-Machine-Comunication

: Groß, M.; Luttermann, H.

Pacific Graphics '93. Proceedings Vol. I
Pacific Conference on Computer Graphics and Applications <1, 1993, Seoul>
Conference Paper
Fraunhofer IGD ()
Bilderkennung; Bildkompression; Bildverarbeitung; Hauptkomponentenanalyse; neuronales Netzwerk

The following paper describes a case study and an evaluation of the usability of principal component analysis for the automatic identification of human faces. It aims at an expansion of an image data set by means of the eigenvectors of its covariance matrix and at an extraction of facial features by projecting face images into a subspace defined by a set of these eigenvectors. This well known image coding scheme has already been introduced as a global approach to face recognition by Kirby and Sirovich and has been adapted by others. Our paper extends the concept by applying more complex nonlinear neural classifiers and compares the results with those obtained by using the simple Euclidian ones. In this way, we provide a carful investigation of the raliabitility of this method for a limited set of persons, as for instance a user group of a computer. For this purpose we built up systematically a data base with entities of different facial expressions and examined both classification and generalization accuracy. First, we introduce the principle component analysis stemming from statistical data analysis and describe the neural network classifiacation concept for it. Then, we compare two classification concepts, namely the Euclidian distance and the backpropagation network. The third part of our paper is dedicated to the investigation of the capabilities and limitations of this method for face recognition. Dur to the fact that this method does not provide an optimal representation of the clusters in featur space, we show in particular, that the use of more complex classifiers improve the accuracy of the method significantly.