Extended Report on Very Low Bit Rate Facial Texture Coding

In this paper we present two methods of enhancing the eigen face approach for the compression of images of human faces. In earlier work, a series of image basis functions (eigen faces) is created using the KarhunenLoeve transform on a large set of face images (the training set). The image to be encoded can then be approximated by a linear combination of these basis functions. We propose a reversible, non-linear geometrical normalization step to be performed on both the training set and the image before it is being encoded. To further increase quality, a partitioning step of the images into blocks can be performed before the KL transform. An improvement on JPEG of 4 dB has been measured on individual images for low bit rates.