Basic Contents

Computer Vision ... (Published 1992)

Simulated Human Vision..... Ian Overington

Location: Eastbourne. UK
ianoverington@simulatedvision.co.uk ............ www.simulatedvision.co.uk

chain, and therefore more inaccessible. Secondly, it has to be admitted that at least some amount
reasoning must be brought to bear to make decisions. Nevertheless, experience with studying
human visual performance can help greatly in deriving algorithms to carry out at least
some of the
processes. In this chapter I would like to consider one
very important aspect of visual interpretation
- that of rudimentary form recognition. In everyday visual tasks an ability to sense the local
characteristics of profiles, such as local curvature, existence of corners, whether the local profile is
smooth or ragged etc., contribute greatly to our ability to recognise objects in the scene. These
observations we carry out with extreme ease and typically within a single glimpse (less than 1/3
second). Now, form sensing by conventional computer vision systems tends to be in terms of
various forms of 'primitives' such as Marr's or Nishihara's groups of cylinders, where such cylinders
are each composed of several pixels. These must be compared with the ability of human beings to
achieve high levels of recognition for images covering only a very few receptors (pixels) on the retina
(Chapter 2), implying a capability to extract form feature data (via the early local processes
discussed in previous chapters) at each individual receptor (pixel) location. A step in the right
direction has been taken by the suggestion to use 'curvature primitives', whilst
a practical method of
curvilinear feature extraction by analysis of
clusters of edge points has been proposed by Suk and
Sull. However, even
these suggestions appear to fall a long way short of known human

12.2. Background
12.2.1. Performance implications.
12.2.2. Recognition criteria.
12.3. Local feature extraction.
12.4. Feature synthesis.
12.5. Practical algorithm implementation.
12.5.1. Curvature and corner measurement.