Non-Geometric Energy Formulation for Adaptive Image Compression

This paper proposes a new adaptive scheme to cope with the limits of the conventional 2D wavelet. The principle is to search for the warped version of the original image which is best adapted to this isotropic kernel. Unlike most prior works, the scheme does not rely on geometry detection or exhaustive search for the best local directions. It formulates a Description Length of the warped Image which does not depend on any a priori about geometry. We demonstrate that its minimization can be compared to a motion estimation between two frames. The warped image can then be efficiently coded by any conventional wavelet coder. An application is presented where the warping is modelized by a regular 2D-mesh. At low rates, results show a significant reduction of ringing artefacts compared to JPEG2K.