Mediation Between Central and Peripheral Processing: Useful Knowledge Structures

Mediation Between Central and Peripheral Processing: Useful Knowledge Structures

Roger Browse

Abstract

This paper outlines a computational vision research project aimed at the development of techniques for the mediation between central and peripheral processes. The key ingredients are structural relations between image and scene domain hierarchies, and a representation which emphasizes the dependencies that exist within the knowledge. These constructs may be used to select areas of the image to process in greater detail on the basis of the progress of interpretation; the nature of the task which is motivating the visual system; and the contents of both peripheral and foveal vision.

Introduction

Information extraction is an essential component of intelligent behaviour. This extraction often involves the selection of sequences of intense, localized processing within contexts of less detailed global processing. The locomotion of an organism through its environment provides locations from which detailed information may be obtained. Of course the extent to which information may be received from locations nearby may vary, but it is always constrained by the capabilities of the organism-1. Similarly the saccadic eye movements of the visual system, together with the acuity structure of the retina, provides a sequence of select locations in the visual environment which may be intensely processed within the context of acuity-limited peripheral vision-2. Visual attention may also be viewed as the selection of locations for more detailed processing, both within the information available in a single fixation of the eyes, and within the knowledge which is involved in the visual processing-3. To extend the idea to its limit, thought may be a sequence of availabilities of information within the mind-4.

Visual attention and thought have no directly observable manifestations as do locomotion and eye movements, and hence the patterns of their operation must be inferred from psychological experimentation and from introspection and will thus always remain speculative.

The interesting questions which arise are:

Is there some uniform computational process which can mediate between detailed, local information extraction and less detailed, global information extraction?
What is the role of such a process in intelligence?

This paper describes part of a research project which is aimed at the development of answers to these questions. The context is a computational vision system which interprets line-drawings of human-like body forms-5 (The drawings are similar to those used to illustrate the movement notation of Eshkol and Wachmann (1958)). Of particular concern is the interaction between interpretation and the selection of locations to process with foveal acuity.