Home

Basic Contents

Computer Vision ... (Published 1992)

Simulated Human Vision..... Ian Overington

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

'sampling matrix'). Typically, as the subtense of the individual raster lines increases from a value
much less than the foveal retinal receptor spacing, the performance first improves, then levels off
and finally falls again. That performance improves with increasing size when the raster line
subtense is small relative to the retinal receptor spacing is no surprise. Under such conditions the
raster modulation is effectively invisible to the eye (due to the blurring by the dioptrics), and so we
have basically a manifestation of visual threshold laws with increasing image size. Equally it is not
unduly surprising that for very large and obvious rasters the performance first limits and then
eventually degrades. Under such viewing conditions eventually all that can be seen is the raster.
Both the small size and large size effects are equally true of images discretely sampled in two
dimensions. What
is, perhaps, surprising - and is of great importance for computer vision studies
using presampled images - is that for raster sampled images it is found that the
best performance is
reached when the spatial pitch of the raster lines subtends around
three retinal receptor spacings
rather than one. Although no fully objective data are to hand,
subjective assessments suggest that
this finding is also equally true for images discretely sampled in two dimensions. The implication of
these findings is that best performance is achieved when the Laplacian-like differencing mechanisms
of the eye are operating on essentially some form of
interpolation of the discretely sampled input
image. Thus it is both unfair and inefficient (compared to human visual performance) to offer
discretely presampled images
directly to any Laplacian-like (or other) differencing operator, without
first considering some form of preprocessing.

Another feature which can be of importance in presampled imagery is response variation from pixel
to pixel. Some form of superimposed pattern is particularly likely to occur on images generated by
2-D discrete sensor arrays (where the individual sensor response almost inevitably produces a form
of sampling noise), on some types of thermal images (where inbalance between individual sensors in
a 1-D strip of elements scanned mechanically results in characteristic striping in the image), and on
some full colour images a variety of registration problems, due to the mechanics of colour video

Continued