Multiresolution image compression using Gaussian Markov random fields

This paper presents a new scheme for image compression that is based on a multiresolution Gaussian Markov random field (GMRF) model. Given an image, compression is achieved by using a resolution-invariant GMRF to parsimoniously capture the spatial correlation present in the image, and then supplementing this information by adding to it a coarse-resolution decomposition of the image. The given image is reconstructed from this information by expressly minimizing the expected mean-squared error (MSE). The algorithm used to obtain the minimum MSE reconstruction has several attractive features: (1) It is non-recursive, (2) It involves only linear operations, (3) It is exactly implementable, and (4) It can provide optimum reconstruction at multiple resolutions. The experimental results obtained by applying the scheme to a variety of images seem to indicate that this methodology of compressing images has potential, and it illustrates the usefulness of MRF models in compressing images