#290925 - 29/11/2006 14:56
Maximum entropy, deconvolution, post-hoc anti-shake
|
carpal tunnel
Registered: 13/07/2000
Posts: 4180
Loc: Cambridge, England
|
The attached is (a crop of) a picture of a co-worker's failed attempt to "get away from it all" by going for a shopping weekend to New York, only to find a giant illuminated billboard of our product directly opposite her hotel. Unfortunately she was so traumatised that the picture has severe camera shake. Now the left and right arrows on the leftmost picture of the Vibez are nearly point sources, so the little quadrants they appear as, depict the convolution function of the whole image. So the question is, is there any software out there that can observe (or even derive) that convolution function and deconvolve the picture, restoring focus?
Back when I was a student there was this technique called "maximum entropy" which essentially iteratively approached the most probable starting image given a convolution function and a final image; it could be used to devastating effect for reading the number-plates of motion-blurred cars, but was (in 1991) on the verge of computational unfeasibility. Has it reached the mainstream yet?
The GIMP plug-in "refocus-it" will do this for out-of-focus or motion-blurred pictures, where the convolution function is either Gaussian or linear, but makes no headway with camera-shake images.
Peter
Attachments
291205-msg.jpg (222 downloads)
|
Top
|
|
|
|
#290926 - 29/11/2006 17:23
Re: Maximum entropy, deconvolution, post-hoc anti-shake
[Re: peter]
|
pooh-bah
Registered: 15/01/2002
Posts: 1866
Loc: Austin
|
tonyc showed me this a few months back. And that's all the help I can provide.
|
Top
|
|
|
|
#290927 - 29/11/2006 20:53
Re: Maximum entropy, deconvolution, post-hoc anti-shake
[Re: RobotCaleb]
|
carpal tunnel
Registered: 30/04/2000
Posts: 3810
|
There was a paper in SIGGRAPH '06 on this exact topic. You can get their MatLab code and may well be able to deblur your picture. EDIT: one of the authors of that paper gave a talk here and asked members of the audience to submit blurry pictures in advance for them to crunch on (this stuff takes hours to run, so they can't do it as a real-time demo). One of the things they *can't* deal with that might be a problem with your picture is whenever you max out a color channel. Once you get a pure white pixel, they don't know whether it's just kinda white or whether it was much, much brighter, which makes it hard to know what to do with the energy captured in those pixels. Looks like you may have some problems along those lines. Most likely, what you'd do is use their fancy deblur thing as a starting point, and then use Photoshop to clean up the output.
Edited by DWallach (30/11/2006 00:37)
|
Top
|
|
|
|
#290928 - 29/11/2006 23:01
Re: Maximum entropy, deconvolution, post-hoc anti-shake
[Re: peter]
|
pooh-bah
Registered: 19/09/2002
Posts: 2494
Loc: East Coast, USA
|
Easiest method = Ask someone in the city to take a cab and resnap the photo? It's still there, right?
But debur / deconvolve would be awesome.
_________________________
- FireFox31 110gig MKIIa (30+80), Eutronix lights, 32 meg stacked RAM, Filener orange gel lens, Greenlights Lit Buttons green set
|
Top
|
|
|
|
#290929 - 30/11/2006 19:35
Re: Maximum entropy, deconvolution, post-hoc anti-shake
[Re: FireFox31]
|
addict
Registered: 14/01/2002
Posts: 510
Loc: NY
|
Easiest method = Ask someone in the city to take a cab and resnap the photo? It's still there, right?
If you're interested in having the photo rather than the method of correcting it, I'm down at the garden regularly.
_________________________
Heather
"I distrust those people who know so well what God wants them to do because I notice it always coincides with their own desires." -Susan B Anthony
|
Top
|
|
|
|
#290930 - 01/12/2006 09:10
Re: Maximum entropy, deconvolution, post-hoc anti-shake
[Re: Heather]
|
carpal tunnel
Registered: 13/07/2000
Posts: 4180
Loc: Cambridge, England
|
Quote: If you're interested in having the photo rather than the method of correcting it, I'm down at the garden regularly.
Trekstor eventually sent us the actual video playing on the billboard. FWIW the text is "The sexiest MP3 player. For only ($199/8GB). Designed in Germany". The video draws attention to the similarity between the Vibez's gentle curves and those of the female (and male) form. Subtle it ain't, but OTOH it must all by itself represent a big fraction of all Rio ever spent advertising their players.
Peter
|
Top
|
|
|
|
#290931 - 01/12/2006 11:04
Re: Maximum entropy, deconvolution, post-hoc anti-shake
[Re: peter]
|
member
Registered: 27/01/2006
Posts: 142
Loc: New Jersey, USA
|
Quote:
Quote: If you're interested in having the photo rather than the method of correcting it, I'm down at the garden regularly.
Trekstor eventually sent us the actual video playing on the billboard. FWIW the text is "The sexiest MP3 player. For only ($199/8GB). Designed in Germany". The video draws attention to the similarity between the Vibez's gentle curves and those of the female (and male) form. Subtle it ain't, but OTOH it must all by itself represent a big fraction of all Rio ever spent advertising their players.
Peter
So sad but so true. I've never owned anything other than Rio players (PMP 300 > 600 > Volt SP250 > Karma) and can't recall ever seeing an ad for any of them in any shape or form.
|
Top
|
|
|
|
#290932 - 01/12/2006 17:26
Re: Maximum entropy, deconvolution, post-hoc anti-shake
[Re: RobotCaleb]
|
carpal tunnel
Registered: 20/12/1999
Posts: 31597
Loc: Seattle, WA
|
Quote: tonyc showed me this a few months back. And that's all the help I can provide.
Holy shit I want that software.
|
Top
|
|
|
|
|
|