Connectionist Pyramid Powered Perceptual Organization: Visual Grouping with Hierarchical Structures of Neural Networks

This paper describes a new approach for organizing image data. Perceptual organization in biological vision is the process of pre-attentively grouping and structuring perceptual information into shapes and forms. In computer vision, processes like these are required in order to transform pixels into structural forms that can be used in higher-level interpretation. Pyramidal structures have been used for perceptual organization in computer vision. The tapering of the structure provides a basis for the \bringing together" of spatially separate image elements. Artiicial neu-ral networks are a means of a system learning by example rather than being pre-coded with the necessary rules. (A problem with performing perceptual organization with computer systems is that biological perceptual organization is not that well understood.) This study focussed on the grouping of line segments into longer lines. A multi-layer feed-forward neural network is successfully trained with backpropagation to perform grouping of line segments within a 2 2 neighbourhood at the bottom of a pyramid. Training at higher levels presents diiculties. The successful training at the base of the pyramid with these small neighbourhoods is encouraging. Better grouping behaviour can be anticipated based on the extra information in larger neighbourhoods.