Re: Scientifically Based Presharpening for Enlargement



Dave Martindale wrote:

"aruzinsky" <aruzinsky@xxxxxxxxxxxxxxxxxxxx> writes:
-------------------------------------------------
So if you did something different from the paper, it's even more important
to tell the reader what you did!
-------------------------------------------------
I suspect most proprietary algorithms are "something different from a
paper." Do you understand "proprietry?"

If your algorithm is proprietary, why not say that?

-------------------------------------------------
Again, you allege something that is completely false. The input image,
without presharpening, clownBox0.25X.png, has absolutely no halos. You
can prove this to yourself by inspecting a nearest neighbor enlargement.
There are halos in the other images from the presharpening and from DDL
and bicubic enlargement.

You're right that the very first image does not have halos. Those were
created by the interpolation in the viewing software I was using
(Irfanview). But the first image *does* have dark-to-light transitions
in the space of a single pixel, which cannot happen with a bandlimited
source. This causes standard interpolation algorithms, which are based on
classical sampling theory, to ring and overshoot when interpolating such
edges. This image is not bandlimited before sampling, as classical
sampling requires.


Classical sampling doesn't require anything. Perfect reconstruction from
the sampled data requires that the data was originally derived from a
bandlimited signal. That was my previous point, you are never going to
be completely successful at recovering the larger high res image unless
you first remove the frequency content that is above what the small low
res version can handle. In reality sampling a bandlimited signal is
rarely (if at all) possible, so investigating how close you can get for
a signal that is not bandlimited certainly seems worthwhile.




And the presharpened image *does* have halos, indicating that the
presharpening is not simply compensating for loss due to finite sensel
area.


I don't understand this business about halos. How does the presence of
halos indicate "presharpening is not simply compensating for loss due to
finite sensel area". The fact that the sensor has area means that all
frequencies that are greater than DC will be attenuated when the image
is captured. To compensate for this all frequencies greater than DC need
to be amplified. This is what a sharpening type filter does for all
frequencies except at (or very near) Fs/2. I don't see what is the
problem with that?


-jim



-------------------------------------------------
I do see differences between the two images, but those differences are
minor compared to the really obvious problems in the images. I'd be
much more interested to see how your method works on high-quality images
without the artifacts. To be blunt, the input images look bad. If they
were mine, I wouldn't show them to anyone. They will look bad after
enlargement as well. That makes it hard to judge the value of the
presharpening.
-------------------------------------------------
Again there are no artifacts in the input image and the simulation closely
matches pictures from a Sigma SD9 digcam with Foveon sensor, except without
noise.

As noted in another message, the SD9 is the worst camera for aliasing
artifacts of all those introduced in recent years.

-------------------------------------------------
It is well recognized thoughout literature that bicubic produces jagged
edges and halos. I used -0.5 in the examples.

When the source is properly bandlimited before sampling?

-------------------------------------------------
And that waviness in the road texture looks completely unrealistic. I'd
rather have a uniform blur than that.
-------------------------------------------------
Again, higher PSNR.

But it's garbage. It shows highly-visible structure that simply was not
present in the original scene. The PSNR simply says that the mean
squared error, on a pixel-by-pixel basis, is smaller. That doesn't mean
the result *looks* more accurate to a human observer, and the latter is
what is important.

The literature has many examples of why PSNR is a really bad way to
measure image distortions, other than noise.

So if you did something different from the paper, it's even more important
to tell the reader what you did!
-------------------------------------------------
Most people who care would guess from the title words "Data Dependent"
that the method is locally adaptive and therefore nonlinear. I see no
overwhelming reason to confirm their suspicions.

Are you aiming your page at professionals in image processing, who are
likely to know this, or random members of the public? Other discussions
of the term "decimation" led me to think you are also aiming at the
non-expert.

Anyway, it's pretty clear by now that you don't really want any input
from anyone who doesn't agree with you, not matter how
well-intentioned.

Dave

----== Posted via Newsfeeds.Com - Unlimited-Unrestricted-Secure Usenet News==----
http://www.newsfeeds.com The #1 Newsgroup Service in the World! 120,000+ Newsgroups
----= East and West-Coast Server Farms - Total Privacy via Encryption =----
.

Quantcast