Architectural Solutions For Intermediate-Level Vision

Abstract

A difficult problem for designers of real-time machine vision systems is handling the iconic/symbolic interface. While image-to-image transformations are readily computed by both pipelined and image-parallel architectures, and list-based and logic-based processors handle symbolic information efficiently, there has all too often been a bottleneck in handling the conversion of iconic data into symbolic form and vice-versa. There is a wide variety of operations that convert images into scalars, contour lists, or other non-iconic descriptions; however, these operations are relatively inefficient on most commercial and research systems.

There are several approaches that can be taken to improve the performance of image-processing systems on iconic-to-symbolic and symbolic-to-iconic operations. One of these is to provide a flexible parallel processing system that can be configured at one time for image-to-image operations, and then for image-to-symbol operations. Another approach is to create special modules that compute particular iconic-to-symbolic transformations such as chain codes from binary images or that compute minima, maxima, means and variances of image intensity data. Yet a third approach takes existing iconic processors and symbolic processors and marries them more tightly than has been done in the past. This article discusses these approaches, focussing on the algorithmic implications of the third one.

Research supported in part by N. S. F. Grant IRI-8605889.