Re: Scientifically Based Presharpening for Enlargement

"aruzinsky" <aruzinsky@xxxxxxxxxxxxxxxxxxxx> writes:
-------------------------------------------------
1. If the main point of the page is implementing a method from a paper,
include a one-paragraph summary of that method in your web page. I'll
use that summary to decide whether I want to read the full paper, and it
will also serve as a substitute if the site with the full paper is
temporarily down."
-------------------------------------------------
I said "loosely based." In fact, the paper has an error that I corrected
and my derivation is very different.

So if you did something different from the paper, it's even more
important to tell the reader what you did!

-------------------------------------------------
2. As someone else pointed out, your examples are horrible. The
discussion talks about how images from a sensor are not point sampled,
they are effectively convolved with a small box (the area of a sensel)
before being measured. So show us an example of an image that is either
*direct from a sensor, with no processing applied*, or at least a
simulation of such an image. (For example, you could take a much higher
resolution image and convolve/downsample it to get your example).
-------------------------------------------------
Why do you assume it is not such a simulation? In fact it is.

Then where do the halos around (for example) the light poles come from?
A raw image from a sensor would not have that. The effect of sensels
that are not points is to attenuate some of the high frequency content
of the image - it would not boost any frequencies. Yet the "before"
images show sharpening artifacts.

-------------------------------------------------
What you have on the page at the moment shows really horrendous sharpening
artifacts (halos) as well as compression artifacts that completely
overwhelm the small difference between the two cases that you are trying
to show. Your starting images should be uncompressed and unsharpened.
-------------------------------------------------
There are no compression artifacts in the input images because PNG uses
lossless compression. The differences are obvious to my eyes. If you
don't see the differences you don't need presharpening.

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.

-------------------------------------------------
Also, if the "data dependent Lanczos" filter is what's responsible for the
weird wave-like textures in the road surface at the bottom of the
image, don't use this filter! Adding artifacts that weren't there in the
original is generally a bad idea, and especially bad in something
that's supposed to be an example. Why not use bicubic polynomial
interpolation, which (a) is familiar to many people, and (b) doesn't
create such artifacts. Yes, it doesn't preserve high frequencies as well
as Lanczos, but you should still be able to see the difference
between the images with and without presharpening applied, and that's the
point of the example.
-------------------------------------------------
a. I provided the input images so people could test them with their
favorite enlargement method.

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.

b. That "weird wave-like texture" is less offensive to the eyes than the
jagged edges you will get with bicubic and the halos will be similar with
both. Providing an exotic example without jagged edges illustrates the
versatility of the method.

Bicubic interpolation does not give me "jagged edges". You *can* get
somewhat "braided" looking effects on high-contrast diagonal edges if
you are using an interpolation method that also boosts high frequencies,
but not all bicubic polynomial interpolation methods do that - it
depends on the coefficient of the cubic term. (If I remember correctly,
a coefficient of -0.5 does not boost high frequencies, one of -0.75
does).

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

Although convolution is linear and commutative, linear operations are not
generally commutative. And, my Data Dependent Lanczos enlargement method
is not linear.

OK. So you should describe it too. The regular Lanczos interpolation
method *is* linear.

Dave
.

Relevant Pages

  • Re: Blind deconvolution???
    ... That's flagrant cheating. ... In your grey example, the artifacts are terrible. ... deconvolution codes. ... images that were measured with finite apertures (where high frequncy ...
    (sci.image.processing)
  • Re: Slightly OT: Photography Archaeological finds
    ... Presentation of artifacts is definitely ... important to archaeologists, and pretty much every archie has had to ... 1)White backgrounds - all the finds have to be shot on white ... I've seen images using clear plastic supports, ...
    (sci.archaeology)
  • Re: Slightly OT: Photography Archaeological finds
    ... Presentation of artifacts is definitely ... important to archaeologists, and pretty much every archie has had to ... awkward positions to photograph and then photoshop it out. ... I've seen images using clear plastic supports, ...
    (sci.archaeology)
  • Re: Slightly OT: Photography Archaeological finds
    ... Presentation of artifacts is definitely ... important to archaeologists, and pretty much every archie has had to ... a white background is not always the best background for every ... I've seen images using clear plastic supports, ...
    (sci.archaeology)
  • Re: pixel artifacts
    ... The two images don't seem to be from the same viewpoint, and it seems that the test application is in yet another viewpoint. ... Other people encountered the same artifacts so I guess that the problem is not related to hardware but to software. ... At the horizon, where the rasterized triangles become very small, you can see white holes. ... These artifacts only appear if I am using vertex shaders to transform the geometry. ...
    (microsoft.public.win32.programmer.directx.graphics)