By Hans Du Buf and Joao Rodrigues
A complete image ontology can be obtained by formalising a top-down meta-language which must address all possibilities, from global message and composition to objects and local surface properties. In computer vision, where one general goal is image understanding, one starts with a bunch of pixels. The latter is a typical example of bottom-up processing, from pixels to objects to layout and gist. Both top-down and bottom-up approaches are possible, but can these be unified? As it turns out, the answer is yes, because our visual system does it all the time. This follows from our progress in developing models of the visual system, and using the models in re-creating an input image in the form of a painting.