Edge-Model Based Representation of Laplacian Subbands

Abstract

Multiresolution subbands have a characteristic structure composed of sparse positive and negetive linear patterns present in close spatial proximity. This implies that there must be efficient ways to represent these subband images as compared to general images. Pixel and block transform approaches based on regular spatial sampling are unale to take this structure into account. This work introduces a novel way of representing Laplacian subbands using (oriented) edge elements. The representation is based on selecting an appropriate set of 7×7 primitives that captures the type of structures present in Laplacians. Unlike contourlets, the primitive set does not constitute a basis but has the twin advantages of small set size and close correspondence between set elements and edge elements that can be interpolated using prior models. As the chosen primitive set is not a basis, the computed representation is formulated by matching given primitives to various image regions, as opposed to decomposing given regions in terms of a basis set. This representation can be used for edge sharpness preserving magnification required in super resolution. The representation can also be exploited for lossy compression and noise removal

Index terms: subband representation, primitive set, image magnification, scale-space interpolation, super-resolution.

This work was funded by DRDO through Proj CAR-008. Authors wish to thank Director CAIR, ISYS-DO and colleagues in CVG for their support.