Microsoft Tech Can Deblur Images Automatically

An anonymous reader writes "At the annual SIGGRAPH show, Microsoft Research showed new technology that can remove the blur from images on your camera or phone using on-board sensors — the same sensors currently added to the iPhone 4. No more blurry low light photos!"

Comments

Enhance

[TheSwampDweller]

Enhance!

Re:Enhance

[bl4nk]

You forgot to include the link to the YouTube video: http://www.youtube.com/watch?v=Vxq9yj2pVWk

Re:Enhance

[slasho81]

Don't forget the TVTropes page: http://tvtropes.org/pmwiki/pmwiki.php/Main/EnhanceButton Sorry.

Re:Enhance

[g2devi]

Actually, that's a Alpha version. The production version is demonstrated here:

http://www.youtube.com/watch?v=KUFkb0d1kbU

Re:lol yea sure

[Helios1182]

Microsoft Research puts out a lot of really interesting and successful research. They aren't the people programming the OS or office applications.

Interestingly simple concept

[Manip]

This is like one of those "Why didn't I think of that?" ideas that you wonder why your camera doesn't already have. The nice part is that it can be done very cheaply (relative to the cost of a camera) and would improve images in many cases. My only tiny little concern is that you might introduce artifacts into your photos - which makes me wonder if it wouldn't be better to store a raw image and the data from these sensors independently? I wonder if there is a scenario where you might be moving but the object you're taking a picture of is stationary relative to your movement. Like for example you're standing on a boat rocking in the waves, you take a photo of the deck, and this technology compensates for the rock which results in a ton of blur.

Re:lol yea sure

[nacturation]

Probably only half-working coming from microsoft

It could be worse... the GIMP developers could have built it, in which case it would be a mostly working implementation of half the features of some existing software. However, nobody would realize this since only the developers would be able to comprehend the UI.

Frankencamera.

[Greger47]

Step back! This is a job for Frankencamera. Run it on your Nokia N900 today.

OTOH having that Arduino board and a mess of wires attached to your camera does score you a lot more geek cred than photographing using an plain old mobile phone.

/greger

Re:Frankencamera.

[slashqwerty]

It's worth noting that page nine of the Frankencamera team's paper mentions the work of Joshi et al when it discusses deblurring pictures. Neel Joshi was the lead researcher from the article we are discussing.

"No more blurry low light photos!"

[Average_Joe_Sixpack]

Social networking sites are about to get a whole lot more ugly

comparison to other methods?

[supernova87a]

I recall that some other cameras, like a Casio I've seen a friend using, also do deblurring, but rather by stacking of rapid subframes (I guess using bright reference points). If I understand correctly, this new method is operated on a single frame. I wonder if anyone has a useful comparison of the hardware requirement/image quality/useability differences between the two methods?

Re:lol yea sure

[binarylarry]

Have you used GIMP in past 5 years?

Re:Okay.

[profplump]

This isn't for people who want to learn photography and take good pictures, it's for people who are shooting their friends in a bar at night to post on Your Face in a Tube and laugh about for a week before being forgotten -- it's merely intended to allow point-and-click shooting work more reliably in poor conditions on cheap equipment with inattentive and untrained operators.

Re:Okay.

[kurokame]

That would be a great point if it involved learning something more complicated than bracing your hand.

MicroSoft is impressive at SIGGRAPH

[peter303]

For the past 8 years or so, MicroSoft has been co-author on more papers than any other organization at SIGGRAPH. This is impressive because SIGGRAPH has a the highest paper rejection rate of any conference I know of - they reject (or downgrade to non-published session) 85% of the paper submissions. And you have to submit publication-ready papers nearly a year in advance, with a video summary.

This reminds me of Xerox PARC - great R & D output, poor commercialization of these results. People wonder if their lab was a toy-of-Bill or a tax write-off.

Useful, but limited

[AliasMarlowe]

It won't help at all if the object is moving. In fact, this feature should be switched off if you're trying to photograph a moving object with the camera (common enough, and not just in sports). It would not be able to compensate for a mismatch between the object speed and your tracking movement, and would do entirely the wrong thing even if you tracked the moving object perfectly for the shot. In this case, there is no substitute for adequate light and/or a fast lens and/or a smooth accurate tracking movement.

As another comment, deconvolution requires a very accurate approximation of the true convolution kernel, which may be provided by the motion sensors. However, to reconstruct the image without artifacts, the true kernel must not approach zero in the Fourier domain below the Nyquist frequency of the intended reconstruction (which is limited by the antialias filter in front of the Bayer mask). In fact, if the kernel's Fourier transform has too small a magnitude at some frequency, the reconstruction at that frequency will be essentially noise, or will be zero if adequate regularization is used. If the motion blur is more than a few pixels, this will generally mean that the reconstructed image will have an abridged spectrum in the direction of blur, compared to directions in which no blur occurred. Of course, if your hand is so shaky and the exposure so long that blur occurs in all directions, then the spectrum of the reconstructed image will be more uniform. It is likely to be truncated compared to the spectrum of an image taken without motion blur.

The quality of the reconstructed image would also be limited by the effects of other convolutions in the optical pathway. For instance, if you're using a cheap superzoom lens, don't expect to get anywhere near the antialias filter's Nyquist frequency in the final image, as the lens will have buggered up the details nonlinearly across the image even before the motion blur is added. If you're using nice lenses (Canon "L" series or Pentax "*" series and suchlike), then this will not be an issue.

The method would seem to be useful in low-ish light photography of stationary objects. A sober photographer would beat a drunk photographer at this, but the technique would help both to some extent. A photographer using a tripod would do best, of course.

Re:Useful, but limited

[AliasMarlowe]

Do you mention FTs just for reference, or are you implying that they are typically used in deconvolutions? In my experience, signals with any amount of noise are much better handled with iterative algorithms.

Yes. Fourier methods are unlikely to be used in practice on images. However, it's instructive to look at the process in a transform space to understand the extent to which information is irrecoverably lost in the optical path. The consequences can be explained using any suitable integral transform, but engineers are most familiar with Fourier and wavelet methods.

Suppose the Fourier transform of the "perfect" image is J, and the Fourier transform of the exact blurring kernel is K, then the transform of the blurred image is L=JK. If the kernel is known exactly, and has adequate magnitude through the frequency range of interest, the image can be recovered simply by using an inverse of the kernel J=L/K. Due to the physics of photon detection, the measured image will also contain some amount of noise N, so even in this ideal case, the recovered image will be corrupted by an amount of noise J'=(L+N)/K. Realistically, the kernel is not known exactly, and may have small magnitude at some frequencies so that its inverse is unreliable. A related consequence is that the blurred image at those bands will be primarily noise. Use of pseudo-inverses is also unreliable, since they are discontinuous near a spectral zero with very high sensitivity to perturbation near that zero. So-called Wiener deconvolution attempts to circumvent this by diagonally biasing the transform of the estimated kernel before inverting, but with generally unsatisfactory results.

Information which has been destroyed cannot be recovered, by any method. Any attempt to do so would merely amplify the noise in the measured blurry image at those bands. Iterative methods (typically a variant of Richardon-Lucy for images) try to minimize the amplification of noise in various ways, all imperfect but preferable to a direct Fourier method. Most iterative methods will, left to themselves, converge on the same asymptote as the Fourier method. However, iterative deconvolution methods always employ a regularization step in each iteration, whose primary purpose is to attenuate adjustments in bands where the kernel is uncertain. They also generally use a small number of iterations, since the first iterations are less affected by noise than later iterations. The end result is that information which was destroyed is not spuriously recreated from noise (in principle at least).

If you're interested, I recommend: P. Jansson (ed.), Deconvolution of Images and Spectra, 2nd ed., Academic Press, 1996. Alas, it appears to be out of print (and my copy is not for sale).

Even so...

[fyngyrz]

Clearly (pun intended) the results have a ways to go yet. Look at the coca-cola image, at the 'a' on the end of the cola... that thing is hosed by the blur, and they're unable to recover it because there's no intermediate contrasting color. Same thing for the spokes on the car rims.

This problem can't be completely solved post-picture. Only large-scale elements with nothing else around them will yield pixel-sharp solutions.

The optimum way to correct blur is to apply active or passive (e.g. tripod) stabilization to the lens prior to the shot; active technology is already pretty decent (photographers tend to measure things in stops; it's intuitive to them... when they say an active stabilizer "gives you" four stops, for instance with Canon, what they mean is that you can shoot four stops slower with the shutter and you won't get blur from camera movement.) Doesn't solve subject movement at all, but then, nothing really does other than cranking down the exposure time.

So... considering lens stabilization has been in-camera for years, and this requires more hardware, but gives you less... I'm going to go out on a limb and say it isn't of interest to camera folks. Maybe in some esoteric role... a spacecraft or something else with a tight power budget where stabilization can't be done for some reason (certainly measurement takes less power than actual stabilization)... but DSLRs and point-and-shoots... no.

Re:Even so...

[EvanED]

I'm going to go out on a limb and say it isn't of interest to camera folks. Maybe in some esoteric role... a spacecraft or something else with a tight power budget where stabilization can't be done for some reason (certainly measurement takes less power than actual stabilization)... but DSLRs and point-and-shoots... no.

Well, sort of, I disagree somewhat. For starters, take camera phones. What do they need to do this? I'm too lazy to read the paper, but seems like accelerometer data. How many phones come with accelerometers nowadays? Pretty much every smart phone does. So no extra hardware there. Second, think of an actual camera. Sure, a lot of P&S cameras now have IS, and a lot of SLR lenses have IS. Maybe that gets you an extra stop or two. But what if you *also* had accelerometer data to apply? If you were in a really low-light scenario and a tripod was impractical (for any number of completely realistic reasons), could this give you yet another stop?

I mean, you say that the results aren't great, and point out flaws. And the results definitely aren't great. Actually, if you look at the Coke image, the whole thing has a very substantial double image -- it looks like the image was translated a couple dozen pixels and added to itself. But in some sense, "does it look as good as a completely stable" is the wrong metric -- if you ask me, the revised images do look rather better than the version before processing. Low-light photography to me always is a huge game of tradeoffs between using a slow shutter and getting blurring, using a fast aperture setting and getting narrow DoF, and using a high ISO setting and getting high noise. And for that reason, I would welcome anything that gives more choices in that arena. (In my dream world Canon stops pushing the pixel count for a couple generations and just works on decreasing noise.)

Now they just need to attach this to Ballmer's

[melted]

Now they just need to attach this to Ballmer's head to deblur the company vision a little.

Bitching about gimp

[fyngyrz]

You know, you -- and 99% of the others bitching about the Gimp -- you're utterly full of shit. I write commercial image processing / editing / animation / generation software for a living, I'm expert - you can read that as "terrifyingly exert" - with Photoshop, Gimp and a whole raft of others... and Gimp is an easy to use powerhouse.

Now I will grant you exactly ONE thing, and that is, you need to sit down and learn to use it. That should take a few hours if you're familiar with something (anything) else; maybe a week hunting down tutorials, or a day hanging with a qualified mentor, if editing bitmaps is all new to you.

If it takes you longer than that, you're either stupid or lazy.

There's *nothing* significantly wrong with the Gimp. It has its limits, like everything does (Photoshop has some really annoying limits too), but for the vast majority of image processing and touch-up needs, it's very nice.

Oh, mommie, my crop function is in a different menu... Some people just need a good smack in the head.

If you really knew what you were doing, you'd have, and use, a whole suite of these programs, because for the big ones, there are areas where they excel, and that's the time to put them into play. If you can't learn to use them because the keystrokes are different, or there is a different paradigm... it isn't the program that sucks. It's you.

Also, if you actually knew how to use them, you wouldn't be bitching about them.

Information théory

[Vapula]

Information théory tell us that once some info has been lost, it can't be recovered. If the picture has been somehow "damaged" by some motion blur, the original picture can't be reconstructed.

On the image, we'll have much more than the motion blur from the camera's movement :
- noise added from sensor electronic noise
- blur from target movement
- distortion coming from lens defect (mostly for low end cameras)
- distortion/blur from bad focus (autofocus in not perfect) ...

The operation that will reduce the camera's motion blur will probably increase the effect from all other defects. You reduce one kind of image destruction and increase the impact of the other one.

Re:Information théory

[beej]

But they are adding information to the system with the additional hardware attachment with all the gyroscopes and so-on. This information can be used to improve the photo, correcting some of the damage. So information wasn't "lost"'; it was just reacquired from a different source, as it were.

It looks like camera shake blur would be reduced, but target motion blur would remain intact.

Of course, if you do a 90-second exposure of the sun, it's likely going to be all-white no matter how much shake-correction occurs. But this solution wasn't meant to fix that problem.

I think you oughta look at the examples.

[Estanislao Mart�nez]

There are some full-size samples of the results of the technique, where you can compare the original image with the result of their technique, and the results of two older techniques. Their technique show some very obvious problems:

1. Doubling of high-contrast edges that are "ghosted" in the original because of the motion blur. In the original, presumably, the motion was something like this: start at position A, hold for a relatively large fraction of the exposure, then quickly move to position B, and hold for another large fraction of the exposure. This means that the photo records two copies of any high-constrast edges, one corresponding to A, and the other to B.

   There are several examples in the link that seem to be like that. The technique doesn't seem to figure this out in all cases, and renders the two ghost lines as separate, sharp lines. Most obvious example: the edge of the front rim of the red car in the second photo. Though compare with the result they got in the photo of the Coca-Cola cans, where it did figure it out for the rack, but not for the text on the cans, and where it introduced some artifact lines perpendicular to the rack.

2. Severe white sharpening halos around edges.

The more instructive comparison is the results of these guys' techniques with the older techniques. Clearly, they're doing a lot better than the older techniques. Still, this is very far away from primetime, IMO.

Re:lol yea sure

[nine-times]

Yeah, Microsoft does some decent research and develops some interesting technologies. It's turning things into products that they seem to have trouble with.