IIRblur does three different types of blur: Exponential Decay,
Resonant Lowpass and Gaussian.
They are all implemented as fast IIR filters, so the
computation time does not depend on the blur size.


Written by Marko Cebokli,  jul 2011  and released under GNU GPL


RELEASE NOTES

** jul 2011
Version 0.1
"pre-alpha" (throw it out and see what happens... :-)





Description of the parameters:

"Amount"
Just that. Controls the effective size of the blur kernel, the
bigger it is, the more the image is blurred.

"Type"
Selects among the three types of blur: Exponential Decay,
Resonant Lowpass and Gaussian.
- Exponential Decay is the fastest. It's equivalent kernel has
a peak in the center, making the effect a little similar to
those "soft focus" lens attachments.
- Resonant Lowpass has a smoothness very similar to a
Gaussian, and is the default choice.
-Gaussian is the slowest of the three, and is included mostly
for completeness sake. It is visually very similar to the
resonant lowpass, so its main use would be in some scientific
application, where a "measurement quality Gaussian" is
required.

"Edge"
Enables edge compensation ("anti vignette"). At the
edges of the image, the convolution kernel "hangs over",
causing some vignetting, which is visible especially with
larger blurs.
This option uses an average of the edge pixels to eliminate
that vignetting.





A short explanation of blur techniques

Image blur is a 2D convolution, which is a rather slow process.
When done directly, the computation time is proportional to the
square of the kernel size. Therefore, people have invented
several techniques to speed it up.

The most popular is the use of a Gaussian kernel, which can be
separated into two 1D kernels. This way, the computation time
is only proportional to the kernel size, not it's square. The
"bokeh" of a Gaussian is very smooth.

The fastest possible blur is a convolution with a flat 
rectangular kernel ("Box blur"). Using the "running average"
algorithm, it runs at just four additions per pixel per color,
regardless of kernel size. However, it's anisotropic sinc type
frequency response produces a very ugly "bokeh".

Last but not least are IIR ("recursive") techniques, as used
in this plugin. Somewhat more computation is needed than for
box blur, but the amount is still independent of kernel size,
and "bokeh" quality is way beyond box blur. Already with only
two taps, a very nice "bokeh" can be achieved.




A note about the functions in the "fibe.h" file

FIBE stands for "Fast IIR Blur Engine". The functions
implement quadrilateral IIR filters (the same IIR filter
is run in four directions).

FIBE-1 uses single delay IIR cells. Their exponential decay
type response gives it a distinctive non-isotropic look,
whose "diamond bokeh" might not be to everybody's liking.
However, it is very fast, consuming only cca five MACs per
pixel per channel (color), regardless of blur size.

FIBE-2 uses two delays IIR cells, which can give a better
circular symmetry. It runs at slightly more than nine MACs
pppc, which is comparable to a 3x3 direct 2D convolution, and
faster than all, except the tiniest size, separable FIR Gaussian
blurs. It can give a very nice, soft and symmetric blur. I
judged it's quality good enough, to merit being the default
choice.

FIBE-3 uses three delays IIR cells, which is enough for the
Young-van Vliet [1] approximation of a Gaussian kernel to
within 0.17% (9 bits). It runs at something more than 13
MACs pppc.

The algorithms are very simple, but the code is a bit bloated,
because I did some optimizations for speed.
The algorithms are so simple, that using SSE didn't bring
much speedup - the limiting factor is memory bandwidth, not
arithmetics. Therefore I decided not to include the SSE versions.



[1]  Ian T. Young, Lucas J. van Vliet:
     Recursive implementation of the Gaussian filter
     Signal Processing 44 (1995) 139-151


