|
 |
Using this software, could I remove all the obnoxious blurring from the
YouTube video? :-P
On 12/10/2011 01:08 AM, Darren New wrote:
> I'd pay for that.
I take it you've never seen blind deconvolution in action before then? ;-)
I notice that the images have a very slightly level of blur to start
with, and none of them appear to be JPEG-compressed. Deconvolution is
not some magic trick; you cannot recover what has been lost from the
signal. You can only reconstruct from what remains. Which means that if
the image is badly blurred, you get much poorer results. And if the
image is JPEG compressed (as all standard camera photographs are), your
results might be very poor indeed.
That said...
http://deconvolve.net/
Post a reply to this message
|
|
|
|
Darren New <dne### [at] san rrcom> wrote:
> I'd pay for that.
> http://www.youtube.com/watch?v=xxjiQoTp864
So the idea is not just to sharpen an image and try to guess what the
original might have looked like, but instead analyze the image and try to
figure out the path that the camera took during the exposure, and then just
"undo" that movement?
--
- Warp
Post a reply to this message
|
|
|
|
On 12/10/2011 04:30 PM, Warp wrote:
> but instead analyze the image and try to
> figure out the path that the camera took during the exposure, and then just
> "undo" that movement?
Basically, yes.
Camera shake is basically a spatial convolution. When you convolve
something in the spatial domain, you multiply it in the frequency
domain. In other words, some [spatial] frequencies are amplified, while
others are attenuated. If you can figure out /exactly/ how the spectrum
was altered, in theory you can apply the reverse alteration, and get
back the original image. Mathematically, that's quite a simple
operation. It's called "deconvolution".
Out here in the Real World, there are several very big problems.
1. How to figure out the exact path of the camera shake, using only the
blurry image? Without knowing the original image [that's kind of the
whole point], it's mathematically impossible to get the "correct"
answer. Instead, you must apply various heuristics.
2. Some frequencies may have been attenuated so much that they get lost
in the noise floor of the signal. If you try to amplify them back up,
you just get signal noise. Other frequencies may have been reduced to
zero amplitude. Now you must /guess/ what the original was. Again,
heuristics.
3. If the image has lossy compression, the "lost" data is probably the
exact information you need in order to unblur the image.
4. The blurring may not be uniform over the entire image. For example,
if the camera rotates, one corner might be near the center of rotation
and hardly blurred at all, while the opposite corner might be severely
blurred. Now calculating the blur just had a whole lot harder. (Let's
not even dwell on how objects at different distances from the camera
move by different amounts if the camera's viewpoint changes.)
5. Any signal noise on top of the blurred image throws the analysis off.
For example, if the lens has dust on it, or there was static on the CCD
or whatever.
Fairly obviously, the worse the blur, the harder it is to unblur. More
extreme blur basically means more frequencies have been filtered out and
have to be amplified / guessed.
I note in passing that this technique also works for image focus as well
as motion blur.
Post a reply to this message
|
|
|
|
It's neat to see deblurring techniques approaching a point where it's
sensible to include them in consumer software. I assume if it this
feature doesn't make its way into Photoshop immediately it'll be because
it's too slow (10+ minutes for a large image I'd think) or because the
parameters would be difficult to tune for non-technical users. Even in
that case, I imagine it'll just be a few more years before these
problems are fixed since Adobe's been quite active in pushing the state
of the art and speedily incorporating those advances into it's products.
On 10/11/2011 5:08 PM, Darren New wrote:
> I'd pay for that.
>
> http://www.youtube.com/watch?v=xxjiQoTp864
>
Post a reply to this message
|
|
|
|
On 10/17/2011 7:13, Kevin Wampler wrote:
> Here's an actual image of one of the examples in the video. It's much easier
> to see the quality of the result here:
>
> http://i.imgur.com/qha1n.jpg
I'd pay for that!
--
Darren New, San Diego CA, USA (PST)
How come I never get only one kudo?
Post a reply to this message
|
|
