A Model of Form Perception and Scene Description

Abstract

This paper examines a computer-programmed model for human recognition and description of scenes of objects that extend over time as well as space. When the scene contains several objects, the model will first output a single most salient name - which may be the name of a high-level whole, like "couple," or of a salient lower-level object, like "stick-girl." It will then describe that salient object, and go on to mention other objects in the scene. The basis of the system is a parallel-serial memory network of interrelated transforms, or characterizers, which are themselves complex sets of probabilistic n-place relations over "things" (where a "thing" can be a value, an internal or external name of an object, part, quality, class, or function, a characterizer to apply, or an act to effect) - a network of simplified neuron-like elements interacting at synapse-like junctions.