Orientation contrast detection in space-variant images

Abstract

In order to appropriately act in a dynamic environment, any biological or artificial agent needs to be able to locate object boundaries and use them to segregate the objects from each other and from the background. Since contrasts in features such as luminance, color, texture, motion and stereo may signal object boundaries, locations of high feature contrast should summon an agent’s attention. In this paper, we present an orientation contrast detection scheme, and show how it can be adapted to work on a cortical data format modeled after the retino-cortical remapping of the visual field in primates. Working on this cortical image is attractive because it yields a high resolution, wide field of view, and a significant data reduction, allowing real-time execution of image processing operations on standard PC hardware. We show how the disadvantages of the cortical image format, namely curvilinear coordinates and the hemispheric divide, can be dealt with by angle correction and filling-in of hemispheric borders.