Using geometrical information for accurate scene understanding in an artificial vision system

The scene understanding stage of an artificial vision system, which is constructed on the principles of the line-based approach of artificial intelligence, is presented. This receives a line-drawing representation of the viewed 3-D scene and produces its structural and depth analysis (e.g., characterization of the surfaces as background or foreground regions and of the edges as terminating, convex or concave parts of the objects, discovery of surface orientation, relative depth and tilt of the edges). The scene understanding stage comprises three processes: (1) Line-drawing encoding - The nontrivial tasks of vertex classification and line discrimination of the line-drawing are performed, whereas the line-drawing is represented in an accurate, compact, and uniform manner. The clockwise order is assigned to the lines connected to each vertex. Encoding employs geometrical information of the lines in the line-drawing and is realized in an elegant (flexible and computationally economic) fashion which can be easily applied to any domain. (2) Initial analysis - Region segmentation, interior/exterior line categorization, discovery of the inside/outside relations of the regions separated along contours of exterior lines as well as region grouping are accomplished employing a combination of the enhancement technique and the boundary stroll. The initial analysis reduces the complexity and ambiguity of the ensuing line-drawing analysis. (3) Line-drawing analysis - A Harmony Theory artificial neural network implementation of a combination of artificial intelligence line-labeling schemes is utilized. This executes parallel constraint propagation and outputs the labels for the lines of the line-drawing that produce the most globally coherent interpretation of the structure and depth relations in the 3-D scene. (C) 1996 John Wiley & Sons, Inc.