Digital BW, The Print

Hi Paul,

Mmmm. My 'understanding' of this was a result of making a 4x4" image of 
black and white lines one pixel wide - so 40x40 pixels alternating black and 
white lines, 10ppi.

If I upsample this by 10% using nearest neighbour interpolation to keep 
things at their simplest, i.e. make it 11 ppi with resample checked, an 
extra white line is inserted every 10 lines, giving a band of white 2 pixels 
wide every 10 lines.

If I upsample by 10% using another method such as bicubic, an extra line is 
still inserted every 10 lines, but some lines on either side are now altered 
in color. This makes it harder to see the extra inserted line, but it can be 
seen. It even seems to show the different effect of bicubic smoother and 
sharper.

BUT, whatever method I use, the end result is banding of varying width and 
sharpness every 10 pixels. How does this square with what you are saying?

Bob Frost.

----- Original Message ----- 
From: "Paul D. DeRocco" <pderocco@...>

That would be strange, but that's not what happens. All pixels are
interpolated. Each pixel in the result maps to some possibly fractional
pixel position in the original, and its value is computed by filtering the
original pixels in that region. Even if an output pixel lines up perfectly
with an input pixel, it is still computed by filtering that input pixel with
its neighbors, so that it won't look different from those pixels that don't
line up. Yes, this does soften the image, by discarding some high spatial
frequencies, but this is often masked by including a bit of high frequency
boost--I think that's what the new "sharpen" option in PS Bicubic is.

> From: Bob Frost
>
> Mmmm. My 'understanding' of this was a result of making a 4x4" image of
> black and white lines one pixel wide - so 40x40 pixels
> alternating black and white lines, 10ppi.
>
> If I upsample this by 10% using nearest neighbour interpolation to keep
> things at their simplest, i.e. make it 11 ppi with resample checked, an
> extra white line is inserted every 10 lines, giving a band of
> white 2 pixels
> wide every 10 lines.
>
> If I upsample by 10% using another method such as bicubic, an
> extra line is
> still inserted every 10 lines, but some lines on either side are
> now altered
> in color. This makes it harder to see the extra inserted line,
> but it can be
> seen. It even seems to show the different effect of bicubic smoother and
> sharper.
>
> BUT, whatever method I use, the end result is banding of varying
> width and
> sharpness every 10 pixels. How does this square with what you are saying?

Sure, if you use Nearest Neighbor. But I'm under the impression that the
various stairstep upsampling actions that people gush about use repeated
iterations of Bicubic.

--

Ciao,               Paul D. DeRocco
Paul                mailto:pderocco@...

Paul,

They do, but as I said using Bicubic simply blurs the bands - they are still 
there. Try it yourself, it only takes a few moments.

Bob Frost.

----- Original Message ----- 
From: "Paul D. DeRocco" <pderocco@...>

Sure, if you use Nearest Neighbor. But I'm under the impression that the
various stairstep upsampling actions that people gush about use repeated
iterations of Bicubic.

> From: Bob Frost
>
> They do, but as I said using Bicubic simply blurs the bands -
> they are still
> there. Try it yourself, it only takes a few moments.

Yes, it does. What's happening is a form of aliasing. Nearest Neighbor makes
no attempt to bandlimit the data, so the aliasing is horrible. Bicubic does
a reasonable amount of filtering of the high spatial frequencies, but not
enough to eliminate all aliasing. Fortunately, real-world images are rarely
as troublesome as test images constructed from alternating black and white
lines. Such test images are useful for seeing what's going on in various
algorithms, though.

--

Ciao,               Paul D. DeRocco
Paul                mailto:pderocco@...

Bob,
I think the test you created is interesting, but not directly useable for
evaluating interpolation routines.
Vertical B/W lines one pixel wide is right at the nyquist limit for spatial
frequency.  Pushing that data through any filter will look rather messy
compared to more “picture” like image data.  For example, a headshot
portrait and willow tree without leaves may “look” better up-sized and
sharpened using different routines.  On the other hand, if the headshot is
of a lion, and you have the whiskers in focus, maybe not. :-)

Best regards,
John Moody

-----Original Message-----
From: DigitalBlackandWhiteThePrint@yahoogroups.com
[mailto:DigitalBlackandWhiteThePrint@yahoogroups.com]On Behalf Of Bob Frost
Sent: Wednesday, September 07, 2005 4:30 PM
To: DigitalBlackandWhiteThePrint@yahoogroups.com
Subject: Re: [Digital BW] Stair interpolation - was Genuine Fractals

Paul,

They do, but as I said using Bicubic simply blurs the bands - they are still
there. Try it yourself, it only takes a few moments.

Bob Frost.

----- Original Message -----
From: "Paul D. DeRocco" <pderocco@...>

Sure, if you use Nearest Neighbor. But I'm under the impression that the
various stairstep upsampling actions that people gush about use repeated
iterations of Bicubic.





[Non-text portions of this message have been removed]

I find that using nearest neighbor followed by enough of a bicubic to 
smooth out the jaggies gives very good results for some images.

of course there is no "right way."

mark