What is Digg?Sponsored by Dragon Age: Origins
Can't get enough Dragon Age: Origins? Play the flash game. view!
DragonAgeJourneys.com - Play the free companion flash game to Dragon Age: Origins.
89 Comments
- kayakto, on 10/12/2007, -1/+42too bad they only wrote an article, i would like to see it in work
- porksoda, on 10/12/2007, -3/+40if you can't see how an algorithm is applied here, you didn't actually "learn" algorithms in high school.
- warmcat, on 10/12/2007, -1/+34Ah this is not generic "deblurring" but an algorithm that attempts to discover the characteristics of "camera shake" and correct for it. Kind of similar to echo cancelling algorithms in audio. As it says in the article the algorithm can't correct for "inadequate depth of field" or motion blur "smearing".
- templest, on 10/12/2007, -3/+31So I guess CSI will actually be a *little* more realistic, now?
- birge, on 10/12/2007, -0/+25@jerigo: motion blur is caused by relative motion of things in the scene. camera shake causes everything to be blurred in exactly the same way, and thus is can be removed.
- rgov, on 10/12/2007, -2/+24Probably mistaking the term for logarithm. Algorithm is just a more computer sciency term for "procedure."
- rolosworld, on 10/12/2007, -4/+25can I buy contact lens with this on?
- CharlesDarwin, on 10/12/2007, -9/+28They should have left that picture blurry!
- borninda818, on 10/12/2007, -2/+20i want that. Photoshop is only good for slight blurs.
- loveandrockets, on 10/12/2007, -2/+18I walked 8 ft away but the picture on the right didn't turn evil. WTF????
- OBKenobi, on 10/12/2007, -1/+16Anyone try this?
http://www.focusmagic.com/ - Urusai, on 10/12/2007, -0/+12Hey, an MIT post-doc said it, so it must be true. You trust Gordon Freeman, don't you?
- rgov, on 10/12/2007, -4/+13The trouble is that blurring is lossy... it's easy to calculate an average (which is what blurring is). Given (x1 + x2 + x3) / 3 = 20, it's impossible to determine what values were used for x1, x2, and x3, but given two of them you can find the third.
The new algorithm essentially (as warmcat wrote) tries to correct motion blur by negating some of its effects, but still cannot magically fix the mathematical impossibility of removing the blur.
I see a lot wrong with this though. You lose depth information, and the trajectory of the camera isn't very easy to reproduce. The distortion would be greated for close objects and less for far away objects (parallax). And if the photo was taken of multiple objects moving at different speeds, forget about it. - bloqmon, on 10/12/2007, -1/+10Well thats your opinion, but this could also have practical uses outside of fixing your shaky pictures. For instance if someone manages to snap a picture, albeit blurry, of someone committing a crime, this algorithm could help law enforcement find the criminal. This is not all about the consumer.
- slothing, on 10/12/2007, -2/+9One way camera shake is different than motion blur is that the blurring is uniform across the entire image.
Motion blur generally means only part (a person for instance) of the image is blurred. Or several parts of the image are blurred, possibly in different directions. - mookieXL, on 10/12/2007, -0/+7Trust me, you DON'T want to see all details on myspace pictures.
- emptyset, on 10/12/2007, -0/+71) Your right in saying that "determine what values were used for x1, x2, and x3" is impossible from only knowing the average of those values.
2) You are wrong in saying this makes deblurring impossible. Remember that x1 also occurs in a number of other equations involving other quantities than x2 and x3 as well. Similarly for x2 and x3. Solving this ensemble of equations is possible under certain assumptions, as this abstract clearly shows. Google "deconvolution" if you're interested in learning more. - bigdoug2005, on 10/12/2007, -2/+8Combine that with IS and you could eliminate the need for a tripod most of the time.
- dreamlayers, on 10/12/2007, -0/+6Focus Magic is pretty good. It can do wonders for out of focus images sometimes and it can fix blur which is caused by motion in a straight line. It's not perfect and it does sometimes produce a certain artificial feel to the images, but I nevertheless recommend it.
- tallfellow23, on 12/29/2008, -0/+6Saw this paper presentation at Siggraph this year, it's a really unique approach, but still leaves some room for improvement. It was neat seeing the convolution kernels they came up with, then comparing them to light streaks in the photos -- they matched almost perfectly. This algorithm is costly to run, however, so you can only specify a certain portion of the image for it to analyze, which could result in output that is less than optimal.
Their presentation was cool too, they injected a little humor, which is good to see since those presentations can be so boring.
On another note, one of the other presentations was on removing motion blur from still images by fluttering the shutter to created a strobed exposure that could be processed and remove an incredible amount of motion blur.
Check it out -- http://www.cfar.umd.edu/~aagrawal/sig06/sig06.html - olddirtycr, on 10/12/2007, -0/+6Well if you read the article,
“This is a research paper,” Fergus notes, “we’re not going to see this next week in Photoshop,” although the MIT post-doc states the algorithm would be ideally applied to the Adobe software program within another year or two. - Kourgin, on 10/12/2007, -0/+5I thought i would be able to buy/download the program through the link, not just get info on it so i can wonder when I'll be able to utilize it
- Dracos, on 10/12/2007, -0/+5It's not clear from the article, but let's say the algorithm can detect the kind of blur it can correct.
Why not license the algorithm to the camera (phone) manufacturers, and have the processing done when the photo is taken so the user never even sees the blurry photo? - tizz66, on 10/12/2007, -0/+4Good thinking.... Bucket-loads of cash!
- fnot, on 10/12/2007, -0/+4Blurring is not lossy and its not an averaging process (it's true that using an averaging filter on an image produces results looking like an out of focus blur).
Nothing this article speaks of is new. Deconvolution is used when you can estimate (or know) the blurring transfer function (which can be approximated by a 2D Gaussian function for example). All these guys have done is to approximate the camera shake with a transfer function and then convolve the inverse with the blurred image -- that's called deconvolution. The results were so-so anyway. I'd like to know how they mapped the camera shake just using one picture.
The hard part is to do a good approximation of the blurring function, otherwise you'll get some nasty side effects when creating the functions inverse. - ericeman, on 10/12/2007, -1/+5The question on everyone's... or at least my mind is how much will this cost and will it be included in the next photoshop....
- inactive, on 10/12/2007, -1/+5cause then people who have camera phones would see a ***** load of cpu cycles everytime they take a picture
- MikeEnIke, on 10/12/2007, -2/+5How is it in any way misleading?
- ddrirc, on 10/12/2007, -1/+4They should automatically do this to every myspace picture submission.
- inactive, on 10/12/2007, -1/+4with the word awesome
- toaplan, on 10/12/2007, -0/+3Rather than having an algorithm detect the direction of the shaking, wouldn't an authored direction/distance sharpening possibly produce a better result? the algorithm might try to oversharpen an image that has blur for other reasons than camera shaking, and a rotational vs a translational shake may also produce different results.
- sgtpinky, on 10/12/2007, -1/+4I'm negatively digging you in retrospect for having such a poor understanding of basic maths.
- mos6507, on 10/12/2007, -1/+4Somebody should use it on the Zapruder film!!
- ylikone, on 10/12/2007, -0/+3Google needs to get this algo and add the button to picasa!
- Metasquares, on 10/12/2007, -0/+2@robbh66:
This is an important step towards making something useful. Underestimating the importance of theory is very shortsighted. - mjpatey, on 10/12/2007, -2/+4Despite your quite sensible ponderings, I think they've done an amazing job of it.
- birge, on 10/12/2007, -0/+2Nice link. Go ahead and mod me down, but the strobed shutter technique aschwo linked to is actually incredibly original, and far more impressive than the article that actually made it to digg, which as I said, is actually pretty unoriginal. (Reading the article even closer, I'm now ready to offer a $2000 wager that I could do better than it.) I'm starting to think that maybe getting my news chosen for me by 50 IT people with more free time than scientific knowledge isn't such a great idea...
- sixlocal, on 10/12/2007, -0/+2Ever try the high pass/hard light method?
Its way better than smart sharpen, sharpen edges, and focusmagic.
It does give you "halo" type effects like in bad HDR images for the extreme cases, but you can get around those with some spot burning. - kently, on 10/12/2007, -0/+2You can get an earlier version here:
http://people.csail.mit.edu/fergus/papers/deblur_fergus.pdf
From the paper:
4 Algorithm
There are two main steps to our approach. First, the blur kernel
is estimated from the input image. The estimation process is per-
formed in a coarse-to-fine fashion in order to avoid local minima.
Second, using the estimated kernel, we apply a standard deconvo-
lution algorithm to estimate the latent (unblurred) image.
The user supplies four inputs to the algorithm: the blurred image
B, a rectangular patch within the blurred image, an upper bound
on the size of the blur kernel (in pixels), and an initial guess as to
orientation of the blur kernel (horizontal or vertical). Details of how
to specify these parameters are given in Section 4.1.2.
Additionally, we require input image B to have been converted to
a linear color space before processing. In our experiments, we ap-
plied inverse gamma-correction1 with γ = 2.2. In order to esti-
mate the expected blur kernel, we combine all the color channels
of the original image within the user specified patch to produce a
grayscale blurred patch P.
...
6 Discussion
We have introduced a method for removing camera shake effects
from photographs. This problem appears highly underconstrained
at first. However, we have shown that by applying natural im-
age priors and advanced statistical techniques, plausible results can
nonetheless be obtained. Such an approach may prove useful in
other computational photography problems.
Most of our effort has focused on kernel estimation, and, visually,
the kernels we estimate seem to match the image camera motion.
The results of our method often contain artifacts; most prominently,
ringing artifacts occur near saturated regions and regions of signif-
icant object motion. We suspect that these artifacts can be blamed
primarily on the non-blind deconvolution step. We believe that
there is significant room for improvement by applying modern sta-
tistical methods to the non-blind deconvolution problem.
=============
The images in the paper look better blurred, IMAO. - Magadass, on 10/12/2007, -0/+2Cool maybe we can use it on those pictures of the Zune!
- hextraordinary, on 10/12/2007, -0/+1How did an advertisement get so many diggs??? :-/
- stanleyfresh, on 10/12/2007, -0/+1what? no bigfoot jokes?
- Bradl3y, on 10/12/2007, -2/+3Are you seriously taking a negative on your inaccurate comment personally? That's what the negative digg is for.
- arborvitae, on 10/12/2007, -0/+1I use 3D deconvolution algos on biological imaging software but it costs like 20k.
- las3rjock, on 10/12/2007, -0/+1For gory technical details, read paper or view the presentation slides at http://people.csail.mit.edu/fergus/research/deblur.html
- sgtpinky, on 10/12/2007, -0/+1Hehe. For sure. Read the testimony of Dr. Bob R. Hunt:
http://jfkassassination.net/russ/jfkinfo2/jfk4/bhunt.htm - DBCubix, on 10/12/2007, -0/+1Reuters already cornered the market on this. haha
- culbeda, on 10/12/2007, -1/+2I have a feeling that IS, VR, OIS, etc would probably interfere with this algorithm as the results are less predictable. I find the largest problem with IS/VR/OIS is that MOST of the lenses are at least an f/4 and usually higher. This leads to a longer shutter speed and IS/VR/OIS do NOTHING for motion blur. so you end up with a great picture of the wall, plants, and the rest of the background, but not a human subject.
For indoor use, I tend to stick with my 50mm f/1.8. Even then, without a flash, motion blur can occur. - sgtpinky, on 10/12/2007, -0/+1I was thinking exactly the same thing. A directed filter where the user specifies distances to certain targets in the photo, and can just drag some sort of widget to adjust for motion blur.
If you look at the original pic, it would be easy for a human to approximate the direction of the camera shake if he had a widget - and these algorithms aren't usually too CPU intensive for a reasonably sized image, so the results could be checked straight away. - stesun, on 10/12/2007, -0/+1I love this, i can't wait until they post the matlab source (if they do). Would be fun to play with it myself.
-
Show 51 - 89 of 89 discussions
© Digg Inc. 2009
— Content created and posted by Digg users is