Slashdot Mirror
Open Source Image De-Noising
GREYCstoration is an open-source tool able to de-noise, inpaint, or resize 2D color images. This is a command-line program developed by the IMAGE team of the GREYC Lab in France and is available for Unix, Mac, and Windows systems under the CeCILL license. The algorithm is based on anisotropic diffusion partial differential equations. These equations are able to smooth an image while preserving its main structures. The demo page presents interesting examples of color image de-noising and reconstruction. This is a serious free alternative to commercial products like Noise Ninja or Neat Image that perform the same kinds of operations. The tool is still a little bit hard to use (command-line based), but I hope the simple C++ API will ease the integration of the algorithm in more user-friendly interfaces. Previous versions of GREYCstoration are already available in Digikam and Krita.
Can it remove the noise commonly used in CAPTCHA images? Will this be the next weapon in the war against spammers?
Another great free tool is called Picture Cooler. It rivals or exceeds Noise Ninja for certain images. Download it here: http://denoiser.shorturl.com/
. size.matter/
But if you want images with less noise, try and buy a camera with a larger sensor. dSLR's have large sensors as do many Fuji cameras including their tiny P&S models. Most sensors on subcompact P&S cameras measure only 5.76x4.29mm (1/2.5"). Many of the smaller cameras by Fuji use a 1/1.8" sensor that measures 7.18x5.32mm.
A nice explanation of noise and sensor size is here: http://www.clarkvision.com/imagedetail/does.pixel
This looks like it's the closest to the image enhancement on the Enterprise (or any other tv show or movie) that i'm going to come. Very cool! Now to pick out crystal clear faces from distant blurry security cameras...
It seems to perform a lot better when dealing with artificially-added noises than real-life images, as if it already knows how to tackle them.
This pyramid photo has basically been 'ruined' after the denoising, I wonder if we added some synthetic noises in the background while leaving the stone face as is, would this app be able to denoise correctly?
Virtual Betting on Facebook for non-geeks.
This tool looks very cool, but today's semi-pro and pro (and even some consumer grade) cameras will store their images in a raw format which preserves 12 bits per color channel at a minimum.
GIMP can't deal with these. Tools such as ufraw can convert them to 8 bits/channel images such as JPEG but don't allow you to actually manipulate the image in its native color resolution.
Linux seriously needs a good image manipulation tool such as the GIMP with 16-bit or even 32-bit per color channel support built-in. This is particularly important for operations like sharpening.
Cinepaint will do it but it's way behind in features compared with GIMP these days.
What's the hold up with GIMP anyway? You'd think its developers would take this kind of issue seriously and would fix the engine to natively do, say, 32 bits per color channel internally.
Use 'slashdot stuff' in the subject line in any email you send me if you want to get past the spam filter.
It'd be useful to have a sample of the filter applied to video for denoising/scaling down, would be much easier to spot how good it actually is compared to other methods. It seems to introduce its own artifacts in the form of those swirly-patterns, would they look natural from frame-to-frame?
This looks like a great app to use for generating more detailed height maps for flight sims. You could zoom in on your height map to the location your aitcraft if flying towards use this tool to create a more detailed height map out of that smaller height map image and then wack that new mesh into your game as you get closer to the ground.
is a pretty good algorithm. We use it for MRI images. The biggest problem is setting the parameters. Fortunately it's an embarrassingly parallel algorithm so I wrote a version that will run realtime on a video card. It's pretty cool to move the slider and watch the noise fade away, move it further and some of the edges blur, further and you start to lose the image.
"The demo images are more than a little impressive."
I disagree. They are overly smoothed and detail is destroyed. They look like the type of thing a noob makes upon discovering video filters. For example, look at the delicate features in the jellyfish or the pig's hair. This samples look more like demonstrations of soften or posterization filters. They should also use real, not artificial, noise.
by taking off my glasses!!
I just took a look at the example images.. The before and after comparisons look very similar to the results you would get from Photoshop's "Smart Blur" tool. So this might be a new way of doing it, but I don't see anything exciting about it. Am I missing something? Or is someone simply making a fuss for the sake of finding a new algorithm with a fancy name?
The lower ISO you can get the more detail you could capture given that other parameters are fixed. Have you ever shot with Velvia ISO 50 film? -- it creates stunning details. I think Fujifilm discontinued it last year or so. In film the lower the ISO the finer the grain. As far as digital is concerned think of ISO as sensitivity of the CCD. You can turn the gain up to ISO 3200 but you will amplify a lot of noise too.
... after those results, I don't want to hear any more bitching about reading yellow license plate numbers off of the reflection of a doorknob reflected in a one-pixel wide eyeball in a black and white security film shown in CSI!
Yeah, looking at the demo pictures I see this is basically replacing noise with blur.
I was thinking about this recently, and I think what we need is a digital camera which can somehow take multiple short exposure shots one after the other and then combine them into a single photo. The algorithm would have to be smart enough to detect movement of both the camera and the scenery in-between frames, so we're talking advanced software, but it does seem possible.
Otherwise, having to choose between underexposed, noisy, and blurry, when shooting telephoto in anything but the brightest of sunlight just doesn't seem right... I guess I could just carry a tripod with me everywhere :)
I keep pressing my ear to my computer, but the pictures don't make any noise. They're actually pretty quiet.
Leave the gun, take the cannoli -- Clemenza, The Godfather
It can also take one image and repaint it in the style of another image, so you can take a black and white photo and a pencil sketch as inputs and end up with your photo rendered using parts of the pencil image which are similar in form.
Another trick it can pull is creating tileable textures from any image. Sometimes the results are a little surprising if you start off with a picture of people at a party but they are totally seamless.
It comes as a GIMP plugin and is easy to use if you are used to the GIMP.
Anything I post is strictly my own thoughts and doesn't necessarily have anything to do with the opinions of IBM.
Fujifilm has announced plans to "bring back" Velvia 50:
? cid=7-7900-8678
:-)
http://www.robgalbraith.com/bins/content_page.asp
It's a new formulation, which they're tentatively calling "Velvia II," so don't write off Velvia 50 just yet
That's because they *removed* the noise, dummy! :]
I think you missed the GP's point
The idea is that you can eliminate blur caused by camera movement by taking many short exposures (high noise because of the short exposure), then align and average them together to eliminate the noise. This will work, but the downside is that it does require computationally intensive image alignment (to remove the camera movement that would have caused the blur in the first place) But that could be done offline.
Me lost me cookie at the disco.
That's the tradeoff with all noise filters. These samples definitely look on par with Noise Ninja, which is saying something.
I think there is a world market for maybe five personal web logs.
If you're experiencing a lot of noise at ISO400, it's time for a new camera.
Also, in response to your later post, what many DSLRs do for long exposures (usually taken at night and with high ISO and experience a lot of noise in the black areas) is to take another exposure immediately after the first one, but with the shutter closed. Then, the camera knows where the sensor noise is and can subtract it from the actual picture.
So if you take an 8 sec. exposure and your camera freezes up for the following 8 sec, you'll know why.
"Avoid employing unlucky people - throw half of the pile of CVs in the bin without reading them." -- David Brent
They are overly smoothed and detail is destroyed. They look like the type of thing a noob makes upon discovering video filters. For example, look at the delicate features in the jellyfish or the pig's hair. This samples look more like demonstrations of soften or posterization filters.
Sure, he's a noob. That DT-MRI of gray matter paths in your brain based on diffusion tensors is purely the stuff of rank amateurs! Bah, next you will tell me free software authors can make a powerful and easy to use image editor. I'm sticking to well known commercial filters found in Paint Shop Pro version 1.0, you know the one that fits on a floppy. Yeah, that's the freedom program.
Oh wait, those other filters are not helping my pig. Perhaps I'm doing something wrong and need to leave that to the pros as well.
They should also use real, not artificial, noise.
Sufficiently advanced noise is indistinguishable from the stuff that comes out of a cheap imaging device, but it's not magic.
Friends don't help friends install M$ junk.
In fact, there already exists software for that purpose: ALE. And it is open-source too!
They should also use real, not artificial, noise.
Check out very impressive clean up of a PDA camera. That's good. Ordinary smoothing filters blur important details, like those in the watch or the baby's ear. How nice that it is already in Digikam, one of the easiest to use photo managers out there.
Friends don't help friends install M$ junk.
This sort of thing is okay for salvaging photos that can't be retaken, but no amount of computer correction can beat a photo taken with the proper camera settings. I'm an above average post-processor, but my favorite photographs usually don't need anything changed at all.
This space intentionally left blank.
Sure, he's a noob. That DT-MRI of gray matter paths in your brain based on diffusion tensors is purely the stuff of rank amateurs!
Being a research scientist doesn't necessarily qualify someone as having a photographer's eye. Nobody's saying the guy couldn't research circles around any of us. What the parent poster said is his de-noise filter is way too aggressive and obscures image detail. That appears to be true, at least judging by the settings he's using for his demo shots.
Sufficiently advanced noise is indistinguishable from the stuff that comes out of a cheap imaging device
Not really true, because the noise that comes out of any imaging device (cheap or otherwise) is not random. It fits a particular profile that's unique to that model of device, or even that particular unit. Advanced photo filtering algorithms (including those used in the in-camera processors that convert raw image data to jpg image files) use that individual profile to filter noise. They're not trying to figure out what's noise and what isn't on the fly, which is at best an imperfect science, and that's being charitable. They have a good idea before they even look at an image what the noise is going to look like, so they do a better job of removing it without sacrificing detail.
The more advanced filters like NeatImage are also almost infinitely configurable in what noise they go after and where, and how aggressive they are. Now, this guy's algorithm seems to be pretty configurable as well, so maybe he just didn't use very good settings himself on most of his image demos, and the algorithm is actually capable of better results. He does seem like he's a better scientist than image-maker so that's entirely possible.
It would be interesting to see what could be done with this if it was given an intuitive GUI and put in the hands of some real photographers. (Yes, even real photographers have to shoot at ISO 800 and above occasionally, and would benefit from noise reduction that actually works without sacrificing detail.)
Oye
If you're referring to Olympus' "Digital Image Stabilization," here's how they describe it:
Digital Image Stabilization Mode uses a high ISO sensitivity and fast shutter speed to enable you to [blah blah blah]
Nothing fancy here about combining multiple exposures and detecting camera/subject movement; just using higher sensitivity than the 50 or 100 that many P&S users are used to, resulting in faster shutter speeds.
Get a better camera and you won't need to clean them up at ISO 400 and sometimes not at ISO 1600. Nikons new entry level DSLR, the D40, can take outstanding photos at ISO 1600 and even boosted to 3200 looks better than most P&S cameras do at ISO 200-400.
Canon makes some terrific DSLRs as well, some with even better high ISO performance, I'm simply more familar with the Nikon line.
I too have a fridge full of film (Velvia, Astia, Provia), however I accept that I'm a dinosaur and proud of it! Modern digital SLRs perform better than 35mm film in practically every respect and challenge medium format in quality. Only with large format is that not true - and large format is something of a niche! Soon of course even this will bow to the digital revolution.
I suspect she's like me, stuck in the past and quite happy to stay there for the time being! There's still something magical about transparency film. The colour reproduction is very special, with a gamut wider than you can sense in either prints or monitors. Although the gamut of modern digital sensors is just as good, there's no way of actually sensing it, as the display devices aren't up to snuff! Wide gamuts make an enormous difference to an image. The colors you see in nature are far more diverse than those that can be reproduced in print or on a computer screen. It's only by actually seeing these things first hand that one can appreciate the difference, prints look strangely grey an lifeless in comparison.
Ah, transparency film!
I'll stop evangelising now - I'm probably preaching to the converted anyway!
Be nice to people on the way up. You will meet them again on your way down!
This sort of thing is okay for salvaging photos that can't be retaken, but no amount of computer correction can beat a photo taken with the proper camera settings. I'm an above average post-processor, but my favorite photographs usually don't need anything changed at all.
I regularly find I want to fix things in post-processing, but noise isn't one of them. I find getting stuff like color balance, constrast, brightness, saturation etc. very hard to get right out directly from the cam, but usually the auto settings keep it inside the sensor's range and they fix well in postprocessing. The one thing you can't fix is ISO. High ISO = noise. Low ISO = blur. If you try to snap a sports player at low ISO you're FUBAR. If you left it at high ISO and try to make HQ landscapes, you're FUBAR. The auto doesn't have a clue except light levels, it has no idea what you're trying to do.
Live today, because you never know what tomorrow brings
The Mac binaries available on the download page are built for PPC macs. If you've got the developer tools installed on an Intel Mac, you'll want to modify the Makefile to change the X11PATH line to:
/Developer/SDKs/MacOSX10.4u.sdk/usr/X11R6
X11PATH =
Then just "make linux"
True confidence comes not from realising you are as good as your peers, but that your peers are as bad as you are.
(I'm not him although I know his work and his ex-supervisor)
;-)
Also consider CImage, by the same author. CImage is a C++ image processing template library (cue to how much C++ sucks compared to the language du jour and/or LISP/Python/Haskell/OCaml, etc
Concerning the inpainting algorithms that many here find impressive, there has been lots of work in this area. One of the seminal works is the paper at ICCV'99 by Efros and Leung. Many CS people will love that one since it is a fairly straightforward extention of the 1948 Markov model proposed by Shannon himself for the automated production of pseudo-english text (i.e. texts that look and sound english but really aren't). The Practice of Programming book by Kernighan and Pike makes use of that algorithm to compare various languages in a fun way.
The Tschumperlé algorithm works on different principles and is much faster, but their particular Markov model shows the impainting problem is not that difficult in practice.
Has ISO rating being abused by digital photography so much that nobody concerns shutter speed, aperture, and lighting anymore?
What fundamentally matters for high motion scene is faster shutter speed. Higher ISO sensitivity makes sure the picture is more easily exposed. Bigger aperture, as well as the scene being well-lit, let more light into the lens, so these two factors also help with exposure.
It is probably best illustrated by shooting a night scene. With dSLR or SLR, you can program long exposure of 2 seconds or more, so you can film at ISO 200 or lower. However, dP&S cameras can't do long exposure, so usually what happens is that the camera adjusts shutter speed to its maximum at ~1/15 (at which point hand shaking can make it blur) but raise the amplification, i.e., ISO rating. As a result, you get a picture where bright areas appear washed out and colorless, compared to dSLR/SLR where the color and details are nicely preserved.
Another extreme is filming high-speed motion, where a common practice is to make a scene or object brightly lit to compensate for faster shutter speed.
What differentiates a photographer and an amateur is that a photographer has all the options and know their trade-offs, but an amateur talks about ISO all the times because, unfortunately, that's the only adjustable metric about dP&S cameras. Even worse, digital ISO rating only has to do with amplification of sensor signal and not actual sensor sensitivity at all. As a result, you lose dynamic range when you amplify more in higher ISO mode.
I'd rather buy a camera phone than point and shoot...
I once had a signature.
This isn't new software, it's just a new release. My wife is a photographer and we've been using this with a Gimp plugin for quite awhile now.
o ration
Gimp plugin can be found here: http://www.haypocalc.com/wiki/Gimp_Plugin_GREYCst
It does ok and can salvage some photos, some, it can't.
Thanks Freshmeat!! ^H^H^H^H^H^H^H^H^H^H Slashdot!!