Towards Continuous Image Representations

Abstract

We propose in this paper a first step towards the creation of continuous, i.e. vectorial, repre-sentations that are useful for image manipulation. Such pixel-free representations have many advantages and are amenable to operations which are difficult or imprecise with pixels. For exam-ple, they can readily be rendered at different resolutions and they are a better choice for ultra-high resolution applications. We explore an approach in which images are decomposed into structural regions that correspond to specified image characteristics. This is done using relaxation labelling. Information taken at different stages during the relaxation is used to extract sub-pixel accurate continuous structural contours. This accuracy is obtained by using snakes as well as the blur present in images (because of the acquisition process). We propose solutions, adapted to our context, to often mentioned problems of snakes, namely initialisation, parameter determination, and instability. The interior of structural regions is represented to allow the rendering of images as close as possible to the original ones. We propose here two schemes, one using a single colour for each region, the second sampling the original image to allow smoothly varying colour in each region.