The University of British Columbia

This paper presents a wavelet based image compression algorithm specifically tailored for facial photographs. The proposed method dramatically reduces memory requirements for facial image databases. The algorithm makes it possible to reach compression rates of 0.25 to 0.1 bpp without compromising the visual quality of the facial features used for identification purposes. This is made possible by utilizing a novel strategy for quantizing the wavelet coefficients, where the spatial content and the frequency distribution of each input image are combined to produce a quantization scheme which is spatially and frequency dependent and different for each image being processed. The main steps of the algorithm are as follows: detection of more and less important spacial areas in the photograph; discrete wavelet transform; space and frequency dependent quantization customized for a specific frequency distribution of each image; and entropy encoding