A Geometrical Approach to Multiresolution Management in the Fusion of Digital Images

Abstract

In most image fusion-based processes, image information are qualified by both numerical activities held by pixels or voxels (data domain) and spatial distribution of these values (spatial domain). Image data are often transformed (registration, multi-scale transform, etc.) early in the fusion process, thus losing a part of both their physical meaning and their numerical accuracy. We propose here a new image fusion scheme in which spatial information are managed apart from image activities, aiming at delaying the alteration of original data sets until the final aggregation/decision step of the process. The global idea is to independently model image information from the data and spatial domains, design fusion operators in both domains, and finally obtain the image aggregation model by combining these operators. Such a process makes it possible to introduce spatial coefficients resulting from spatial fusion into advanced aggregation models at the final step. The fusion in the spatial domain is based on discrete geometrical models of the images. It consists in applying a computational geometry algorithm stemming from the study of the classical digital coordinates changing problem, and modified to be efficient even on large 3D images. Two applications of the fusion process are proposed in the field of medical image analysis, for brain image synthesis and activity quantification, mainly destined to the automated diagnosis of Parkinsonian syndromes.