Upscaling - Photography

[Bart van der Wolf]

"Hans-Georg Michna" <hans-georgNoEmailPleasemichna.com> wrote in message
news:u0hvgvcg5h59lfg745moc8u5epmsjmmibt4ax.com...

> Many thanks for the good hints!
>
> Does anybody here understand how the upscaling algorithms like
> Lanczos interpolation work? Any experiences?

Do a Google search for a message with a description that Dave Martindale
gave in this forum (much clearer than I could explain it) some months (?)
ago.
"Lanczos windowed sinc" interpolation is very good in retaining the original
pixel contrast when looked at from an appropriate distance. It can cause
some visible (ringing) artifacts on sharp edges if looked at from too close.
Other interpolation methods (such as bi-cubic) can cause loss of pixel
contrast which also looks less than stellar, so it depends on image content
what method is better for a give output size. Grainy/noisy images may even
benefit from some mushifying.
Programs like Qimage can also use a different "Vector interpolation" method
that works very good for many types of images.

Bart

[Bart van der Wolf]

"Hans-Georg Michna" <hans-georgNoEmailPleasemichna.com> wrote in message
news:t4n0hvcpr5re8dp2ubfq12s15gtbq6aki74ax.com...

> "Bart van der Wolf" <bvdwolfnospam.nl> wrote:

SNIP

> >Programs like Qimage can also use a different "Vector interpolation"

method

> >that works very good for many types of images.

>
> Bart,
>
> I'm not actually looking for any kind of interpolation. What I
> want is a program that detects high-contrast edges and redraws
> them sharply in the higher resolution, which would amount to a
> kind of anti-aliasing function.

Try the Qimage Vector interpolation, it "invents" plausible edges, without
the posterized look of Genuine Fractals (and at a better price).

Bart

[Hans-Georg Michna]

"Bart van der Wolf" <bvdwolfnospam.nl> wrote:

>"Hans-Georg Michna" <hans-georgNoEmailPleasemichna.com> wrote in message
>news:t4n0hvcpr5re8dp2ubfq12s15gtbq6aki74ax.com.. .

>> "Bart van der Wolf" <bvdwolfnospam.nl> wrote:

>> >Programs like Qimage can also use a different "Vector interpolation" method
>> >that works very good for many types of images.

>> I'm not actually looking for any kind of interpolation. What I
>> want is a program that detects high-contrast edges and redraws
>> them sharply in the higher resolution, which would amount to a
>> kind of anti-aliasing function.

>Try the Qimage Vector interpolation, it "invents" plausible edges, without
>the posterized look of Genuine Fractals (and at a better price).

Bart,

thanks, will look. "Invents plausible edges" sounds exactly like
what I'm hoping for.

[Time passes.]

Just looking at their web site. Are you sure Qimage does vector
interpolation? The site only mentions Lanczos interpolation, not
a word about anything else. They even have examples that show a
very smooth interpolation, but no edge sharpness.

Hans-Georg

--
No mail, please.

[Don Stauffer]

When you say redraws it in 'higher resolution', that is implying a
resampling or interpolation. Sharpening per se does not necessarily
lead to higher res. It CAN, if the actual resolution of original image
is well below Nyquist limit, but not if it is at limit.

Most sharpening filters are the equivalent of a differentiation added to
a proportional function.

Hans-Georg Michna wrote:

>
> "Bart van der Wolf" <bvdwolfnospam.nl> wrote:
>

> >"Hans-Georg Michna" <hans-georgNoEmailPleasemichna.com> wrote in message
> >news:u0hvgvcg5h59lfg745moc8u5epmsjmmibt4ax.com.. .

>

> >> Does anybody here understand how the upscaling algorithms like
> >> Lanczos interpolation work? Any experiences?

>

> >Do a Google search for a message with a description that Dave Martindale
> >gave in this forum (much clearer than I could explain it) some months (?)
> >ago.
> >"Lanczos windowed sinc" interpolation is very good in retaining the original
> >pixel contrast when looked at from an appropriate distance. It can cause
> >some visible (ringing) artifacts on sharp edges if looked at from too close.
> >Other interpolation methods (such as bi-cubic) can cause loss of pixel
> >contrast which also looks less than stellar, so it depends on image content
> >what method is better for a give output size. Grainy/noisy images may even
> >benefit from some mushifying.
> >Programs like Qimage can also use a different "Vector interpolation" method
> >that works very good for many types of images.

>
> Bart,
>
> I'm not actually looking for any kind of interpolation. What I
> want is a program that detects high-contrast edges and redraws
> them sharply in the higher resolution, which would amount to a
> kind of anti-aliasing function.
>
> Hans-Georg
>
> --
> No mail, please.

--
Don Stauffer in Minnesota
[email]staufferusfamily.net[/email]
webpage- [url]http://www.usfamily.net/web/stauffer[/url]

[Bart van der Wolf]

"Hans-Georg Michna" <hans-georgNoEmailPleasemichna.com> wrote in message
news:osl2hvgjjgj39rpu6rt1rgrciaiifaipti4ax.com...
SNIP

> Just looking at their web site. Are you sure Qimage does vector
> interpolation?

Yes, I use it (almost) all the time:
[url]http://www.ddisoftware.com/qimage/qimagehlp/interp.htm[/url]

This is a page from *before* Vector interpolation was introduced:
[url]http://www.ddisoftware.com/qimage/quality/[/url]

This is an example (on one type of image):
[url]http://www.ddisoftware.com/testpics/l-vs-v.jpg[/url]
and of another type:
[url]http://www.ddisoftware.com/qimage/vector.jpg[/url]

Bart

[Hans-Georg Michna]

"Bart van der Wolf" <bvdwolfnospam.nl> wrote:

>This is an example (on one type of image):
>[url]http://www.ddisoftware.com/testpics/l-vs-v.jpg[/url]

Bart,

thanks, I had missed those at first.

But this isn't what I expected. It is quite good, but I don't
understand why the algorithm makes a perfectly sharp image
unsharp and fuzzy. It shouldn't do this, I think.

The desired result would be an arrow that looks exactly like the
first one, except without the steps.

Hans-Georg

--
No mail, please.

[Hans-Georg Michna]

Don Stauffer <staufferusfamily.net> wrote:

>When you say redraws it in 'higher resolution', that is implying a
>resampling or interpolation. Sharpening per se does not necessarily
>lead to higher res. It CAN, if the actual resolution of original image
>is well below Nyquist limit, but not if it is at limit.
>
>Most sharpening filters are the equivalent of a differentiation added to
>a proportional function.

Don,

I know, but I'm not actually looking for sharpening. I'm looking
for anti-aliasing, taking the pixel steps out when scaling up.

It seems though that the function I'm looking for does not exist
yet as a readily available piece of software.

Let me demonstrate the function with just black and white
pixels.

Picture before upscaling (4 symbols = 1 pixel):

##
##
####
####
######
######
########
########

Picture after upscaling with factor 2 (1 symbol = 1 pixel):

#
##
###
####
#####
######
#######
########

It would need something like edge detection and edge redrawing.

I think the video upscaling chips (Faroudja etc.) do exactly
this. I can see it every day in my home cinema. When the picture
moves rapidly, the chip is too slow, but as soon as the picture
becomes less dynamic, the edges are smoothed.

Hans-Georg

--
No mail, please.

[Bart van der Wolf]

"Hans-Georg Michna" <hans-georgNoEmailPleasemichna.com> wrote in message
news:cjq4hvgdhgpgq5nf846u24mjtcrg0l46uj4ax.com...

> "Bart van der Wolf" <bvdwolfnospam.nl> wrote:
>

> >This is an example (on one type of image):
> >[url]http://www.ddisoftware.com/testpics/l-vs-v.jpg[/url]

>
> Bart,
>
> thanks, I had missed those at first.

No, it wasn't mentioned on the web site, but in the Yahoo support forum, so
you couldn't know.

> But this isn't what I expected. It is quite good, but I don't
> understand why the algorithm makes a perfectly sharp image
> unsharp and fuzzy. It shouldn't do this, I think.

Because the program has to guess what is image detail, and what is a pixel
boundary.

> The desired result would be an arrow that looks exactly like the
> first one, except without the steps.

If the first one would have been rotated by just a few degrees, all edges
would have pixel steps (aliasing), whereas Vector interpolation would have
made a (better) compromise. Then some unsharp masking would produce a
visually superior result.

Bart

[Hans-Georg Michna]

"Bart van der Wolf" <bvdwolfnospam.nl> wrote:

>"Hans-Georg Michna" <hans-georgNoEmailPleasemichna.com> wrote in message
>news:cjq4hvgdhgpgq5nf846u24mjtcrg0l46uj4ax.com.. .

>> But this isn't what I expected. It is quite good, but I don't
>> understand why the algorithm makes a perfectly sharp image
>> unsharp and fuzzy. It shouldn't do this, I think.

>Because the program has to guess what is image detail, and what is a pixel
>boundary.

>> The desired result would be an arrow that looks exactly like the
>> first one, except without the steps.

>If the first one would have been rotated by just a few degrees, all edges
>would have pixel steps (aliasing), whereas Vector interpolation would have
>made a (better) compromise. Then some unsharp masking would produce a
>visually superior result.

Bart,

I guess it's a matter of taste. The original picture didn't have
any unsharp masking (or anti-aliasing using intermediate
pixels). Why should the upscaled one have it?

But this discussion may not make much sense now. Nobody will
listen to me and write a new program. :-)

Hans-Georg

--
No mail, please.

[Hans-Georg Michna]

"Bart van der Wolf" <bvdwolfnospam.nl> wrote:

>In this, and only situations like it (exact integer multiple upscaling of
>sharp edges), a so called bi-linear interpolation would have done just that.
>It demonstrates that the image content (and scale factor) determine which
>interpolation method does best.

Bart,

are you sure? I would have thought it would produce some grey
pixels.

But anyway, I would like an algorithm that repaints edges,
perhaps even if they were slightly unsharp in the original.

Hans-Georg

--
No mail, please.