I am affiliated with AutoQuant, and my views and statements below are
affected by this affiliation.  I am responding with a few comments.

1. Regarding the statement from Wes Wallace about
 It still takes about 8 hours a file of about 300x300x60.

AutoDeblur comes with 5 speed/performance settings where faster speeds are
achieved with an understandable trade in image performance.  A setting of 2,
which is 2nd from the best for robustness gives still very remarkable
deconvolutions, sacrifices very little in image quality for a factor of 4
speed increase.  I.e., 8 hours becomes 2.  Other fast algorithms are
available, including a fast xz slice deconvolver that runs in a few seconds.

2. Regarding Steve Niemela s comment about:
 This ignores the downside issues with the blind approach.  It generally is
less accurate than the PSF approach, it takes longer and it doesn t work
with some kinds of images.

It is unclear to me what is meant by  less accurate  and I am wondering if
this claim could be substantiated.  Our experience is contrary to this
statement.  Our experience, furthermore, is that we can blindly deconvolve
any type of image that is a candidate for iterative and constrained
deconvolution.  Not having the PSF does not add any limitation to the types
of samples that can be deconvolved.  I think that if a sample happens to be
 un-deconvolve-able  , then it is un-deconvolve-able by any iterative and
constrained method.

3. Regarding Steve Niemela s comment about:
  In many instances, the results of our theoretical PSF were better, both
quantitatively and qualitatively, than with the AutoDeblur algorithm.....

How many instances, and what percentage?  I could make a similar claim to
the contrary as we have experimented with the Jansson-van Cittert algorithm,
which I understand that VayTek uses.
I am unaware of the over-smoothing that Steve refers to.  Perhaps there was
something set up incorrectly.  I wonder if I could obtain the data set used
so we can verify the problem.

4. Regarding Steve Niemela s comment about comparing Nearest Neighbors
images side by side with nearest-neighbors and iterative constrained and
blind deconvolution, he said  if you place these images side-by-side you can
see the differences, but they are not big differences.

This is contrary to our experience.  It makes sense that the Nearest
Neighbors method SHOULD give large differences compared to any iterative
constrained method.  First, it is fundamentally an unsharp mask operation (a
well-established image enhancement operator), which is fundamentally an edge
enhancing sharpening filter.  With AutoDeblur, which has both Nearest
Neighbors and blind deconvolution, while the Nearest Neighbors algorithm
gives remarkable results and deserves credit for this in its own right, it
still gives results that are clearly and obviously different from the blind
deconvolution. For example, compared to blind deconvolution, it sharpens
some out-of-focus remnants that are in a plane where they do not belong, but
the blind deconvolution removes these remnants as you would hope.  Again,
this does not take any credit away from the Nearest Neighbor method.  It is
a good method, fast (a few seconds or less) and useful, but interpret images
carefully and if you want those out-of-focus remnants removed it is better
to use an iterative constrained or blind deconvolution.

Timothy J. Holmes, D.Sc.
President, CEO, AutoQuant Imaging, Inc.
Adjunct Assoc. Prof., Rensselaer Polytechnic
Institute
518-276-2138
518-276-3069 fax
www.aqi.com
[log in to unmask]