🚀 go-pugleaf

In article <Bz2n91.J3M@ie.utoronto.ca> green@ie.utoronto.ca (Marc Green) writes:
>psilsbee@ccwf.cc.utexas.edu (Peter Silsbee) writes:
>
>>After all, if the claim you are making is "this method produces superior
>>images in the sense that human viewers find them more visually pleasing,"
>>any numerical measure you choose simply won't back up that claim, and any

Sorry about the sloppy phrasing, by numerical measure I meant "image-
derived numerical measure," in the sense of something that could be
automatically derived. I hope that wasn't the cause of this small
controversy!

>>reviewer who asks for such a meaure hasn't _understood_ your claim (all
>>too frequent an occurrence, I suspect). If, on the other hand, your claim
>>is that "this method produces superior images in the sense that there is
>>less error," then you need to back that statement up with numbers.

again, I was sloppy. Sorry. I think we are in agreement on the point that
human psychophysical judgment is a valid and scientific image quality 
measure; I tried to say that in my first post but apparently was not
very clear about it. Having said that, I do have some quibbles:

>
>This is exactly backward: the only valid, empirical measure of image
>quality is human psychophysical judgment. 

Of course not. It is utterly application-dependent. Some images are never
seen by humans. 

>People outside of psychophysics
>don't seem to understand that human judgments are just as objective and
>"scientific" as any other experimentally derived data. 

(unnecessarily caustic [but with humorous intent] remark about people 
_inside_ psychophysics suppressed, with difficulty)

Well, I may have been more flippant about it, but in essence, this is what
I was saying. That's why I specified that it is important to know what the
authors are claiming about the results of their algorithm. This measure
is appropriate if the claim is that (to quote myself) "this method produces 
superior images in the sense that human viewers find them more visually 
pleasing," which is certainly not the only claim that is ever made about
image processing algorithms.

>It is the "objective"
>measures which are arbitrary and ad hoc. Many studies have shown that 
>mean squared error and similar measures are very poor predictors of
>perceived quality. And since images are created for humans to view, the
>only empirically valid measure of quality is what human observers say it
>is. 

Yes, certainly, IF your images are created solely for humans to view, then 
that is the best measure. Even so, it must be used with care. Perhaps
one reason that it is distrusted is that it, unlike more "objective"
measures, can be manipulated. I can lossily compress and reconstruct
an image, and design a forced-choice experiment that will show that people
can't tell the difference. If my application involves magnifying the image
by a large factor to locate tiny details, then I've been dishonest. 

>There is a long history of attempts to find physical image measures which
>correlate well with human judgments. At the moment, there aren't any. The
>problem is that the mapping from physical images to perception is
>highly complex. To quantify perceived quality, you need to know both
>the physical description of the image and the observers psychophysical
>response. No measure based soley on an image description will ever be 
>much good, unless it reflects the operation of the underlying
>psychophysical mechanisms..

And there are additional complications, such as scale (see above example),
display equipment (for a standard measure, one would need an image
presentation standard, I suppose), etc. And when you start talking about
image sequences, there are even more variables.

--Peter