[Digikam-devel] [Bug 127374] unsharp mask: max. radius way too small!

[Peter Heckert]

Hello,

While I am not a specialist I have done a lot of investigation about 
gaussian filters and
graphics blur- and sharpening filters in common.
So I feel I should comment some of your questions.

Matthias Wieser schrieb:
> What does this mean?
> It seems to be impossible now to use a radius below 1 pixel.
>   
Yes, that depends on the algorithm used. Real Gauss filters would have 
infinite resolution.
This is impossible, because the pixel data has finite resolution. 
therefore gaussian filters
with radius < 10 pixels always arent real gauss filters, but an rough 
approximation for
the theoretic gauss filter.
Some do this approximation in pixel granularity, others try to do an 
approximate
sub-pixel interpolation, but they all are inaccurate at low radius.

Also they all differ in accuracy. However there is only one application 
where high
accuracy is required: Edge detection.
For the simple purpose to blur an image, or to sharpen an image I would 
prefer less
accurate, but faster algorithms, for example multi pass boxfilters.

Thats the other point: All gaussian filters as implemented in gimp, 
cinepaint or
digikam are way too slow for a 100 pixel radius blur. Gimp would need 
1/2 hour
for a multi-megapixel image, photoshop needs 1-2 seconds ;-)

I have written some investigation and little rants about this: ;-)
<http://www.hphsite.de/dcamnoise/theory.html>
<http://www.hphsite.de/dcamnoise/Gaussian-blur.html>

I believe, it would be the best to use simple boxfilters for gaussian 
blur and for
USM, because they are very fast, regardless of radius size (if well 
programmed)
and because they are absolutely numerical stable at low radius, and 
because an
isotropy error of some percent wouldnt be noticeable for this type of 
application.


> Btw.: Is it intentional that most plugin dialogs do not state which units 
> are used?
>   
Normally the standart deviation in pixel units is meant for "radius".
It is really difficult -if not impossible- to calcultate this for low 
radius, and the
exact definition of radius doesnt exist.

> Sharpness: 0 - 100: Percent? Some other unit?
>   
Gaussian blur: 0 - 100: Radius in Pixel?
> Noise reduction: ?? (How do I reduce the strength of the default noise 
> reduction to e.g. 50%? No idea, maybe I should look into the handbook...)
>   
There is no adjustment for 50% noise reduction.
The amount of noise reduction depends from the noise granularity and 
from the
images gamma value and from other imponderables.
The user must find the best threshold level by trial and error.

I try to overcome this, however this would require to process the whole 
image
twice, there is a lot of work to do.
Currently I believe, that the noise reduction algorithm would be better 
used in
gimp or some other image editing application, where the regions of 
interest can
be masked.

In Noise reduction i implemented a gaussian blur recipe taken from here:
<http://www.ph.tn.tudelft.nl/~lucas/publications/1995/SP95TYLV/SP95TYLV.pdf>
(There you can see how complicated this stuff can be ;-)

Recently I have found out, that this algorithm is fine for edge 
detection, because it
has good isotropy at large radius, but it is bad for sharpening, because 
it is
numerically unstable at low radius.
<http://groups.google.de/group/comp.graphics.algorithms/browse_thread/thread/abda3a7841349a52/350b81531d7f1512>

Here are the impulse responses of a gauss-iir filter:

Example:

Radius 0.25:
55    -0.000000
54    0.000001
53    -0.000046
52    0.001762
51    -0.055922
50    1.108410
49    -0.055922
48    0.001762
47    -0.000046
46    0.000001
45    -0.000000
44    0.000000

(It's interesting that this has 2nd derivative as a component.)

Radius 0.5:
56    0.000001
55    0.000007
54    0.000086
53    0.001093
52    0.012896
51    0.122505
50    0.726826
49    0.122505
48    0.012896
47    0.001093
46    0.000086
45    0.000007
44    0.000001


So I switched to boxfilters in my new version.

> Unsharp mask:
> Radius: 0 - 100: Pixel?
> Amount: 0 - 1: Strength normalized to 1?
> Threshold: 0 - 1: Pixel color normalized to 1? But the histogram shows pixel 
> colors in the range from 0 - 255 (for 8 bit images). Should be the same, 
> should'nt it?
>   
Most of them try to mimic photoshop and use arbitrary estimated values.
If you want to know exactly, look into the source.
This will get worse, when it comes to 16-bit images. Is the treshold 
meant as percent,
is it meant as a 0-255 intensity value or is it meant as a 0-65536 
intensity value?
(Even photoshop isnt clear here and also the curves tool misbehaves when 
the image
has gamma=1 and 16 bit)


greetings,

Peter