Newsgroups: sci.image.processing
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!gatech!swrinde!howland.reston.ans.net!news.sprintlink.net!noc.netcom.net!netcom.com!perry
From: perry@netcom.com (Perry West)
Subject: Re: subpixel
Message-ID: <perryD8n8yn.I3H@netcom.com>
Organization: NETCOM On-line Communication Services (408 261-4700 guest)
X-Newsreader: TIN [version 1.2 PL1]
References: <3p8cl4$od7@dub-news-svc-3.compuserve.com>
Date: Mon, 15 May 1995 23:51:59 GMT
Lines: 21
Sender: perry@netcom11.netcom.com

It is possible to make measurements to a precision that is a fraction of 
a pixel.  That is the normal use of sub-pixel resolution.  However, one 
of the criteria for this is that the object must be fairly large.

As far as detecting a feature that is smaller than a pixel.  It is not 
generally possible.  I know of only one condition where it was possible 
to detect (reliably) features smaller than a pixel.  This was an opaque 
part which was intensely backlit so that even a minute pinhole would 
produce a response.

A theoretical limit to *reliably* detecting a feature is to have two 
pixels spanning the feature in its narrowest direction.  In practice, 
four pixels is much better.  In a noisey environment, even more pixels 
may need to span the feature to reliably detect it.  This does not mean 
that you will *never* detect a feature as small as a pixel, or even 
smaller; only that you can't rely on detecting it.

Now, if you want to discriminate between these features, they will need 
to be even larger.  Because we don't know the contrast and noise 
conditions in advance, this is usually evaluated empirically.

