Re: Sub-pixel accuracy for line detection.


"X t l a n" <x t l a n @ i n t e r i a . p l> wrote in message
news:e27f4o$jpk$1@xxxxxxxxxxxxxxx
Hello everybody.

Recently I've been dealing with algorithms for sub-pixel line extraction.
One of them is available at:
http://www.lgi2p.ema.fr/~montesin/PublicationsPS/scia97_subpixel.ps.gz

Could you tell me please, how you understand sub-pixel line extraction and
what particular meaning is carried by the "shifting coefficient" for
applied above Gaussian filter pulse response? Is that also a sub-pixel
location of pixel in an image? I understand reason why filter is shifted
but I also can imagine situation when a 1px width line is captured by a
camera and fitted just between two columns of pixels with the following
coordinates: x and x+1. If brightness of the line is equal to 255 (all its
pixels) and background amounts to 0, all pixels in columns (described by
coordinates x and x+1) will amount to 127 and 127 (the most likely). In
that case each of x and x+1 pixel column will get a half of energy carried
by brightness of the line (in fact: blurred line). On the assumption that
the line is a part of natural picture (e.g. forest), how can I claim that
this is in fact 1px width line and its brightness amounts to 255 while it
could be 2px width line acquired with its actual brightness around 127?
Any ideas, how should I understand a sub-pixel accuracy term?


Regards,
Peter.

Hi Peter,

You can't tell the difference between the two cases you presented. Subpixel
algorithms always need an edge model to figure out a subpixel position. The
whole idea is that the image of an edge has some known grayscale
distribution across several pixels (two, at a minimum) and that subpixel
positions can be calculated by "fitting" the observed grayscale distribution
to the model.

In some cases you might be able to derive your edge model from other parts
of the image. Say you have an edge running across a large part of the image
and you know the edge should be of constant width and grayscale then perhaps
you can form a model based on the width at the point with the maximum
intensity.

Andrew


.

Relevant Pages

  • Re: How do I get decent fonts on Linux
    ... > Err, no, it was sub-pixel anti-aliasing, almost all kerning and character ... > placement will be sub-pixel on any computer depending on your zoom level. ... > pixels were too visible due to the double-height pixels. ...
    (uk.comp.os.linux)
  • Re: How do I get decent fonts on Linux
    ... character placement will be sub-pixel on any computer depending on ... double-height pixels. ... there is an equal but opposite expert ...
    (uk.comp.os.linux)