Multiscale Face Recognition

Good face detection is an important part of a reliable face recognition system. We propose methods of face detection and normalization. Normalization requires the exact location of eyes to straighten any rotated face image. Eyes detection methods are also developed for this purpose. The accuracy of the classical PCA technique for face recognition can be improved by incorporating the class information. We develop a system based on the method of least-squares which takes into account the class information. The discrete wavelet transform decomposes an image into bands (subband images). The importance of these bands vary with different image properties like image resolution. There exists a particular set of subband images that are essential for good face recognition. We construct a model to find these subbands for input images. This model is a function of image resolution. Moreover, we develop a multiscale face recognition system which utilizes the model to find important subbands. These subbands are used for recognition in our system.