Hi Chris
>
>could you please briefly explain how 'Blind Deconvolution' works?
>

Although you asked Wes I like to throw in my 5 cents as well.

With blind deconvolution the object and the PSF is estimated alternatly.
That means that with some initial assumptions of the PSF the first estimate
of the object is being made and in the next step with that 'better' estimate
of the object a new estimate of the PSF is generated and so on.
It is well known that only with noisy observations from your microscope a
blind deconvolution is not feasible. It is necessary to implement a-priori
knowledge. That is only available on the PSF. So, the constraints imposed on
the PSF during blind deconvolution are usually simple ones, like
bandlimitedness and symmetry, etc. These however must come from the optical
parameters of the lens and in a simple way represent therefore a theoretical
PSF in itself. Since these constraints can be described usually as ad-hoc
and hardly rigorous mathematically, it is hard to describe what actually
happens. Despite many factors that can't be modeled (e.g. scattering)
experience show that reconstructed PSF's are indeed reasonable good in
comparison to measured ones.
Although personally I don't see that as a big problem - but the necessity to
determine the PSF could limit the wide usage of non-blind deconvolution
systems. Users want to concentrate on their biological experiment rather
then on esoteric factors. But then the computational effort is much higher
compared to nonblind methods.
On the contrary though, it is not that difficult to acquire a PSF! There is
software available on the market (e.g. Zeiss KS 400) that will basically
allow beads much larger than the PSF itself - utilizing a constrained
iterative restoration algorithm to exploite the circular shape and size of
that bead to restore a PSF out of it. That is equivalent as to restore a PSF
with blind deconvolution on the image data, except here we have knowledge
about the object (sphere).
One other aspect that is worth mentioning is that if blind deconvolution is
carried out in small blocks, the variations of the PSF mainly in axial
dimension is taken into account as well.

I hope that answered your questions
Lutz
______________________________________
Lutz Schaefer
Advanced Imaging Methodology Consultation
16-715 Doon Village Rd.
Kitchener, Ontario
N2P 2A2, Canada
Email: [log in to unmask]
Phone, FAX: (519)-894-8870
______________________________________


-----Original Message-----
From: Christian Lohr <[log in to unmask]>
Newsgroups: bit.listserv.confocal
To: [log in to unmask] <[log in to unmask]>
Date: Friday, March 24, 2000 10:35 AM
Subject: Re: varieties of deconvolution algorithms


>Hi Wes,
>
>could you please briefly explain how 'Blind Deconvolution' works?
>
>Thanks
>Chris
>
>
>
>Wes Wallace wrote:
>>
>> To clear up some of the recommendations surroundint deconvolution
>> software, I think the following information should be noted.
>>
>> There are four kinds of deconvolution algorithms.
>>
>> 1. "Nearest Neighbor" - will run very fast on even the junkiest computer,
>> however, does not result in much improvement, and furthermore, does not
>> preserve brightness ratios.
>>
>> 2. Empirical Point-spread Function - the standard deconvolution software
>> that most packages offer.  This is potentially the highest-resolution
>> method but if the empirical point-spread function is not acquired under
>> incredibly rigorous conditions, the result is worse than nothing.
>>
>> 3. "Calculated Point-spread Function" - this type of algorithm
>> is usually offered along with packages whose main selling point is
>> empirical point-spread function.  here instead of acquiring an
>> empirical psf, the software calculates a point-spread function based on
>> the optical specs of your microscope (i.e. numerical aperture, wavelength
>> of light).  the result is OK, but inferior to blind deconvolution.
>>
>> 4. Blind deconvolution.  This is only offered by Autoquant in their
>> Autodeblur package.  It can be considered the 'sport utility vehicle' of
>> deconvolution.  Potentially, an empirical point-spread function can give
>> higher resolution -- just as potentially, a Ferrari can outperform a
Jeep.
>> However, in a real situation you are probably going to have more freedom
>> and versatility using the Jeep.
>
>--
>Christian Lohr, Ph.D.
>ARL Division of Neurobiology
>University of Arizona
>PO Box 210077
>Tucson, AZ 85721-0077
>
>Phone: (520) 621-6671
>FAX: (520) 621-8282
>[log in to unmask]