Newsgroups: comp.graphics,comp.graphics.algorithms,sci.image.processing,comp.compression
Path: cantaloupe.srv.cs.cmu.edu!rochester!udel!gatech!howland.reston.ans.net!nntp.crl.com!pacbell.com!nntp-hub2.barrnet.net!news3.near.net!analog.com!analog.com!hooper
From: william.hooper@analog.com
Subject: Re: Measuring Disortion
Message-ID: <1995May25.141419.24467@analog.com>
Originator: hooper@zeus
Sender: usenet@analog.com (News pseudo-user usenet)
Organization: Analog Devices, Inc.
References: <3pvs0g$k3n@info-server.surrey.ac.uk>
Date: Thu, 25 May 1995 14:14:19 GMT
Lines: 31
Xref: glinda.oz.cs.cmu.edu comp.graphics:77159 comp.graphics.algorithms:17214 sci.image.processing:14844 comp.compression:20242

In article <3pvs0g$k3n@info-server.surrey.ac.uk> ee11sw@surrey.ac.uk (Mr Scott Walsh) writes:
>I will take the orignal and the corrupted image and then by doing edge
>detection on BOTH of the images, and then finding the difference, or RMS
>error, use this as a measure of the error as this will tel me how much the
>edges in the image, and hence the objects in the image have been distorted.
>
>Well over to you, is this a good or bad idea? Dont flame me too hard, as I
>am a bit of new boy to this image compress/processing lark *8-)

Well, one problem with this method, which is also a problem with PSNR, is that
quantization errors can be easier to see in areas with low energy.  For example,
in a picture with a large amount of blue sky, the sky is not a uniform color but
is instead a slow fade.  When you compress this type of image by quantizing over
the whole picture (or if you are trying to display this image with a limited colormap)
the sky can end up with a "stair-step" look.  This is _very_ noticable but does not
produce a large PSNR and would probably be seen by your method as no error.

I am definately not an expert, but it seems that people use PSNR more as a way to see
how much energy was lost in the compression process, not how much "image quality" was
maintained.  IMHO, the only way to really compare different compression methods is with 
a blind test with a panel.  One other thing -- the method you use to measure loss should
probably depend on what you need from the image.  For a security application where
you would like to be able to do edge processing on an image, the method you propose
would make more sense than PSNR.  For direct to TV applications, what really matters
is how noticeable artifacts are to a person who is not looking for them.

-- 
--Will Hooper               |"Even old New York was once New Amsterdam. |
  Analog Devices, Inc.      | Why'd they change it?  I can't say --     |
  Austin Design Center      | people just liked it better that way."    |
  william.hooper@analog.com |   TMBG                                    |
