A low level vision hierarchy based on an irregularly sampled retina

Biological vision systems process signals extracted by a retina that reduces the dimensionality of visual information for processing by higher cortical areas. To date researchers in machine vision have not reported computing a biologically inspired artificial retina that can sample visual information without over-sampling the central foveal region or creating discontinuities in the retinal tessellation. We have implemented a spacevariant retina that has a uniform receptive field density in the central foveal region and becomes increasingly sparse in the surrounding periphery. The retina contains isotropic receptive fields and the responses from the retina were processed by machinery motivated by biological cortical circuitry. We introduce a cortical graph that enables us to compute spacevariant cortical filters that can sample irregular image information generated by the retina and thereby extract edge information and process colour opponent information. In this paper we briefly describe the selforganising methodology we used to generate retinal tessellations and detail the construction of visual information processing layers in our architecture.