Radon-Hermite analysis applied to image coding

In this paper it is shown how local differential structures in images can be described and coded efficiently by means of the analysis of local projections. Such analysis involves a 1-D Hermite transform of all local Radon projections, where localization is achieved by applying a Gaussian window to the image. Since the Gaussian window is isotropic, the 1-D analysis can be based on the 2-D cartesian Hermite transform. For the case of oriented patterns such as edges and lines, the image is suitably described with a unique projection giving the maximum directional energy. We show that the inverse Radon transform is not required in the synthesis process, since only the information along one direction is preserved. A more specific representation is introduced in an edge model whose parameters are determined in terms of the Hermitian coefficients up to second order. Experimental results are also included.