Early image representation using regions defined by maximum gradient paths between singular points

摘要

A representation for images is described that is intended to be interposed between the pixel data and Marr's primal sketch. The image is segmented into regions named slope districts which are defined by maximum gradient paths between singular points. A rich data structure is formed containing the regions, their boundary components, overall attributes, and relationships with their neighbours. This representation provides a framework to simplify the interaction between bottom-up and top-down processes. Examples are given using simple grouping rules to construct more complex elements like those in the primal sketch.