Integrating Fourier descriptors and PCA with neural networks for face recognition

Abstract

A new approach to the face recognition problem is presented through combining Fourier descriptors with principal component analysis (PCA) and neural networks. Here the faces are vertically oriented frontal view with scaling, orientation, expression, and illumination changes. There are many research activities on face recognition using the face space which is described by a set of eigenfaces. Each face is efficiently represented by its projection onto the space expanded by the eigenfaces and has a new descriptor. Previous work on eigenface has shown that it performs well only with changes in expression, but results are poor in the case of rotating, or scaling the input face. In order to enhance the performance of the eigenfaces technique to accommodate other variations of the input face, the Fourier vector of each face is projected in the eigenspace. Neural networks are used to recognize the face through learning the correct classification of these new descriptors. A real-time system has been created which combines the face detection and recognition techniques. A recognition rate of 91% has been achieved over real tests. It is also shown that our proposed system behaves accurately in the case of rotated or scaled faces as well as for changes in expression.