Computational analysis of early visual mechanisms.

Abstract

An important class of mechanism, image filtering, is normally used to model the first stages in the human visual process. A formal examination of the general computational properties of image filtering, looking at the logic of what image filtering should accomplish and how filtered images can be described, shows that image filtering does not make any useful information explicit. To do this, a further stage of post-filtering processing, termed image description, is necessary. With these image descriptions it is possible to establish relationships between the structures that emerge and the filters used. Three properties emerge that are of benefit: the structures are simply shaped when filters are used that are modestly orientation tuned; the responses are suitably primitive; and the response structures show spatial clusterings that can be used to identify certain classes of object. Within this framework, the properties of a range of oriented filters were examined in a series of computational experiments. Given a particular form of image description, the optimal filters in terms of response structure simplicity, primitiveness and flexibility show a degree of orientation selectivity corresponding to that obtained in mammalian vision.