Inductive generation of icon trees in foveated multi-resolution recognition

Iconic approaches to object recognition rely upon the direct comparison between an unknown input image with of characteristic views of images of a target subject contained in an icon database. Hence, this approach to recognition is distinguished by making comparisons in icon-space as opposed to making comparisons upon a more condensed symbolic representation of the information contained within the database and unknown images. The appeal of the iconic approach is in its algorithmic simplicity, which is traded for increased memory utilisation. However, to be tractable some method of interpolating between the sampled views of the subject is usually necessary. It is also possible to constrain the visual search-space by adopting transformations that confer same form of viewpoint invariance. The RC (retino-cortical) transform has long been used in this capacity and confers both rotational and scale invariance by converting these degrees of freedom into orthogonal linear shifts in the output cortical space. This work is a progression of a RC transform based system that employs both multi-resolution search to improve the efficiency of matching unknown images to those in the icon database. The adopted strategy also attempts to cluster similar views together at each branch in the tree-structured, icon database. The novel aspect of the work reported is the fully automatic construction of the icon database trees using an approach based on dynamic rule induction