Find the complete article in English here: ZERO NOISE PHOTOGRAPHY

I have tested a technique to completely eliminate noise* on digital images based on the signal/noise ratio improvement achieved through overexposition.
At the same time this technique extremely expands the dynamic range of your image in the shadows (don't think of HDR, it's not like that) and recovers in high detail all textures present in the darkest areas of your image.

* It actually does not eliminate noise at all, just takes for every pixel that one with the best signal to noise ratio. That is why textures are not only 100% preserved, but improved.

To do this you simply need to shoot twice making use of a tripod. One shot will be as usual, keeping highlights unburnt. The second shot with be done with a severe overexposition (I found +4EV to be a good value). A simple piece of software merges those two shoots into one final image with no noise on it and fine detail even in the darkest zones. I have converted my modest 350D in a virtually noise-free digital camera with 12 f-stops of real usable dynamic range.

The development of the technique and results are explained in detail in Spanish here: TECHNIQUE OF THE 4 f-STOPS. If I have feedback from the forum I will take some time on a proper translation of the whole article. Meanwhile find here some of the results:

Original shot (the one we would have usually taken, caring of highlights):




Second shot (+4EV with respect to previous one):




Merging result:




Apparently there is no difference: same tones, bright, contrast. In fact that is what I intended to achieve; unlike HDR programs this method provides a new image with the same appearance as the original one. But let's look closer: first let's analyse the 16-bit histograms:

Original image:


Resulting image:


The quality improvement is very noticeable. Histogram is soft and has no zero values as the original had. Non interpolated levels are much closer thanks to overexposition and that means a lot of tonal richness in the definition of textures.

Let's look now at the improvement in some noise samples in the darkest areas (left=original, right=result):








Where usually there was no more than noise, we have recovered perfectly textures that can now be treated using curves or whatever method you like best. Please note that the original shot was done in ISO100 on a 350D; the reason for such a amount of visible noise is no other than these areas of the image are REALLY very dark; I applied the same curve both to the original and processed images to be able to the check the noise improvement.

This is not an all-purpose technique; first of all we need two expositions what requieres some camera stabilization system (tripod). But I think it can improve a lot quality of results in some applications such as:

- Architecture and interior design
- Studio still life
- Landscapes with dark areas
- Night or low lighting conditions photography
- High contrast escenes
- General photography with zone processing in mind

I plan to develop a program to merge N RAW images (not just 2) with arbitrary expositions (not just 0 and +4EV) to automate the process.

Regards.

That looks great ... sounds simple too with just two shots to make the image... I can't believe the final outcome ... both of the shots that you took would have had a certain amount of noise.. but the detail in the final image looks great.. almost if you had been using a fill light. I would certainly use this for my workflow ... what type of program did you make? Is it similar to a lens cast correction program?

Ros

I wrote a routine that checks pixel levels in the overexposed image. If the pixel is burnt, I takes it from the other image. If it is not, exposition is corrected and that low noise level is used. It works!

Some more examples:




Looking at dark areas:

If you have a Mac version of your utility, I'd be glad to do some testing for you.

Regards,
Bernard

GLuijk,
In what way would you say your method is better than standard blending procedures as outlined in this Luminous Landscape tutorial?

http://www.luminous-landscape.com/tutorial...-blending.shtml

Another way to think about this (sort of upside down) is in terms of the expose-to-the-right method discussed here previously. The +4 EV is ETTR. So far, in fact, that the highlights and upper midtones get burned out -- no surprises there. The other image (which has no highlights burnt out) is used to fill in the burnt-out data.

I have wondered for a long time why film scanners did not use this technique to achieve extremely high dynamic range (or low shadow noise). They clearly consider this an important specification, since they all inflate it so much in their marketing Scanners have no stability problem, so aligning multiple exposures should be easy. This technique would allow a scanner to deliver a merged image with full 16-bit depth, even if the sensor is capable of only 12-bit depth each pass. Apparently a few scanners are now starting to do this ....

I'd certainly like to know more about it. Please do the translation. Thanks, Jim

Yes, please do. I'm very interested in this.

Cheers,
Paul

Good question. I have been looking at them, 3 techniques are explained.

First my approach is to provide the same image as the original (no exposition change, no bright change, no contrast change), so that the user can freely use the recovered information in the shades in the way he likes best, but starting from a someway virgin image. Just no noise on it, but expanded dynamic range (I never liked the feeling of HDR techniques).

Secondly, it can work from RAW files, and not only 2 but several of them with any exposition values. With 2 images 12-stops dynamic range is achieved, but with 3 shots 16-stops real dynamic range can be truly achieved with no change in contrast. Allowing change of contrast even more than 16-stops could be obtained, but I think this is too much!

Third is totally automatic, just click a button and wait about 10s for 2 8Mpx RAWs. The result is a 16-bit TIFF file.

Fourth, I plan to offer it for free. If a friend of mine doesn't get to modify the DCRAW source code to perform this algorithm straight into the RAW files before developing them (which would be absolutely fantastic) I will write an application myself. It will read the RAW files but will develop them prior to processing, which is a certain disadvantage compared to my friend's approach.

In fact I already offered a beta simple version for download that takes 2 RAW files with any exposition difference among them and combines into a noise-free output image. Check it here: PROGRAM ZERO NOISE DOWNLOAD
The program is simple simple, but It's a bit tricky. You will need some Spanish assistance as it works only reading the files from certain folder (c:\ceroruidio\) and needs a DLL to be copied into the PC. I offer there my two dining room RAW files for download and test.

Can you show us a sample image of a region spanning the bright-dark transition, where your program switches image source pixels? A good example would be a crop including both the bright lamp and the nearby dark speaker from your first posted image. If the program has any glitches, that is where they will be found.

sure, here it is:



Black areas are taken from the original image while cyan ones are from the +4EV. The election is done pixel by pixel independently, this is the stronghold of automating this method through software as the idea is not new at all.

Another example from a forum member who tested the program:



On the left side of:

This result can already be achieved in PS pretty easily using the "blend if" sliders under blending options. Expose one shot for the highlights and one for the shadows, bring the over exposed shot down in the RAW processor so the tonalities are the same, layer the frames in PS and use the blend if tool to reveal the overexposed shot in the shadow regions only to get noise free shadows. As mentioned before the difficult/problematic areas are where there is sensor blooming and fringing.

Does your program offer any advantages to this or is it a similar idea with automation?

tim

Can't this be done in CS 3 Extended by taking multiple exposures on a tripod and merging under Automate? However it won't extend the dynamic range.

Yeah I pointed that possibility in a forum, as I almost have no idea about PS: select pixels according to their level so those beyond a threshold would be deleted showing the lower layer with the normally exposed image. Could you please try it with my original RAW files? I would be very interested.

NORMAL
OVEREXPOSED

If you can select individual pixels, then I can only see one more problem: RAW developer are usually non-linear, that means that gamma correction and conversion into a space colour are prior to edition, so as the exposition correction (-4EV) to be applied. I wonder if this can reduce perfomance and quality.
A possible solution would be linerar developing like that achieved with DCRAW (in fact my program calls DCRAW from its code).

Best

I must correct this: we trend to call "dynamic range" to those results achieved with HDR software. OK, let's forget for a moment about those tone mapping tools.

Noise is the biggest limitation to record properly detail in the shadows. In this link: 350D dynamic range I tested my 350D on a high contrast image and calculated (aprox.) it has a useable dynamic range of 8 f-stops. In the darker shadows noise becomes so noticeable that the 4 lower f-stops that theoretically can be recorded on a 12-bit RAW are unuseable.

But If you clean the noise in those 4 lower f-stops (as this technique does), you are in fact expanding the real dynamic range your camera can record from a scene. I always refer as dynamic range as that of the scene being recorded, not that of my TIFF file. In fact any image with level values from 0 to 65535 should be called as having 16 f-stops of dynamic range, but it is fiction as does not correspond to real scene dynamic range.

If that CS3 option cleans up noise, you can by certain say you are getting more dynamic range.

Check this histogram representation, it's cool: it represents the same escene but shot at 1 f-stop intervals, and representing the true log histogram. As we can see there are 8 f-stops that record properly the scene before noise makes it unusable:



On the right the white wall I was shooting at (it appears green as I developed without applying the white balance that would scale all levels and thus displace the 3 channels).

This technique has been around since...a long time. Though it is good to see it get some airing again.

As well as blending two stacked layers you can also blend multiple bracketed exposures in photoshop using the merge to HDR function (which in CS3 also has the benefit of aligning the layers as well if there is small movement). You have the tools on the Mac already. NB in this case it is not about dramatically expanding the captured dynamic range and more about increasing the quality of what is captured.

The biggest advantage of this method is a reduction in quantisation noise in the shadows resulting from low (12-bit) conversion of analogue to digital signal. This gives much smoother tonality in the shadows and is useful for those who like to manipulate there images. The reduction in shot and thermal noise comes from averaging over multiple exposures (as per traditional correlation processes to remove noise in signal processing).

yeah but in most medium cameras, quantization problems in the lowest f-stops really don't matter as noise make them absolutely useless.

0EV: 2048 levels, 2048..4095
-1EV: 1024 levels, 1024..2047
-2EV: 512 levels, 512..1023
-3EV: 256 levels, 256..511
-4EV: 128 levels, 128..255
-5EV: 64 levels, 64..127
-6EV: 32 levels, 32..63
-7EV: 16 levels, 16..31
-8EV: 8 levels, 8..15
-9EV: 4 levels, 4..7
-10EV: 2 levels, 2..3
-11EV: 1 level, 1

Last four f-stops are flooded with noise as for instance in the -10EV stop, a noise level of 2 would pull any pixel out of that f-stop.
With this technique no averaging is done, just pixel selection. The equivalent (in the 12-bit range) levels table achieved with a +4EV overexposed image is something like:

0EV: 2048 levels, 2048..4095
-1EV: 1024 levels, 1024..2047
-2EV: 512 levels, 512..1023
-3EV: 256 levels, 256..511
-4EV: 2048 levels, 2048..4095
-5EV: 1024 levels, 1024..2047
-6EV: 512 levels, 512..1023
-7EV: 256 levels, 256..511
-8EV: 128 levels, 128..255
-9EV: 64 levels, 64..127
-10EV: 32 levels, 32..63
-11EV: 16 levels, 16..31


In green are marked those f-stops obtained from the overexposed image. Even in the 12th stop there are levels enough to allow heavy edition. And in addition to this the SNR has improved by a factor of 2^4=16, what makes noise on the 350D practically invisible.

And more than 12 f-stops of real dynamic range is truly possible, thanks to the 12-bit to 16-bit conversion. In fact with an additional +8EV overexposed shot, I am sure that 16 f-stops of dynamic range can be recorded. The difficult part is to find in real life such a huge contrast.

The emphasized text is the difficult part, at least for the test images provided by Guillermo. At least my meagre skills aren't quite up to matching them well enough to avoid posterization effects.

Mostly with non-linear data, though, which is more unwieldy.

The best place to do most image math is in the RAW linear state where everything is very simple.

Despite the disclaimer regarding HDR, I'm not sure that the method is that much different from HDR, which is usually used to gain dynamic range, but can also be used to improve shadow detail as explained by Sean McHugh on his Cambridge in Colour website. See the Tip towards the end of the discussion.



Of course, if you are using a tripod, the first thing to try would be to shoot at base ISO and see if that gave decent shadows. If not, then the technique might be helpful, but I'm not sure if HDR techniques already available couldn't also be used.

Anyway, a good idea worth some discussion.

Bill

The advantage of this method outlined by GLuijk might be the avoidance of halos around bright areas in the scene that one often gets when blending images using layer masks. I'm thinking of situations such as a bright window in a dark room. It's difficult to get an even lighting right to the edge of the window frame.

A method that replaces pixel values rather than averaging them seems likely to solve this problem.

Although I am envious of your mathematical and programming skills, I tinkered a bit with your samples and it seems to me that nearly the same effect can be obtained with CS3 HDR. I also experimented with Photomatix (plug-in) on the CS3 HDR.

This has been a good lesson for me. I tend to be a bit heavy handed with Photomatix and HDR (I'm probably not the only one) and this experiment has reminded me that a "non-HDR" image can be produced using the HDR process.

When the goal is a "normal" photo, the noise reduction using "HDR" is very impressive.

I don't wish to discourage you in the development of what may very well become a new product; however, you must keep in mind that Photoshop and HDR is a competitor.

By the way, I really like the image of the junked bicycles in window light.

Thanks for stimulating a new way of thinking and processing (for me anyway).

Hi all, you are all right: what I am doing is not far from HDR (or at least from the previous steps any HDR soft performs prior to the tone mapping stage), or to any blending technique of 2 or more images into one as CS3 can do. In fact I posted this method in the Spanish forum calling it "soft HDR". The result of combining different images will, among others, be always an increase in dynamic range through overall signal to noise ratio improvement.

What I try to show you (no commercial interest as I don't sell my little programs) is a simple soft that is focused on providing:
- Minimum noise (none of the techniques above can improve it; the more can be expected is to match results).
- Maximum texture detail in deep shadows (there is no blending in fact, just pixel selection which is optimum).
- Natural dynamic range expansion in the shadows end (not tone mapping or so steps applied than transfigure our picture's appearance and give it an unreal appearance).

The goal is provide the cleanest possible image to the user with just a click keeping the original image parameters (bright, contrast and tones). It is now up to him to choose the way how to use it; it could even be used as a free of noise input to any HDR soft (Photomatix for instance can deal with one single RAW, or you can use two versions of it for the PS HDR module), or most commonly be processed using zones without surprises when pumping up the darkest areas.

A friend of mine is modifying DCRAW's C source code to perform all these operations, not only in linear as I do, but over the RAW file itself prior to Bayer demosaicing, white balance or any scaling (in fact we are having some trouble with the black point offset most cameras keep in their RAW files that must be substracted before being able to consider a linear behaviour of the sensor, which is not strictly linear due to this offset).
And if he manages, I'll try to convince him to pursue a 16-bit DNG RAW file as an ouptput. That would be simply great, can you imagine? put a bunch of RAW files, with different arbitrary exposition, into a 16-bit RAW file free of noise ready for developing on your favourite RAW developer.
But I have a feeling that recreating the DNG RAW format is not a joke so perhaps we must be happy with just putting our fingers into the image before the developing process.

Regards.

Fair point - though I was thinking more of the HDR processing which does deal with linear data. It is much harder to blend images after they have a tone curve applied (which is what you are referring to )

If you could manage that, than it would be gigantic. I really would love to use such a thing.

Now if you can't do that it would use to RAW images to produce a Tiff, right ? Now do we than lose all the things from raw like whitebalance, and other adjustments ? That would be a drawback for less noise.

Yes, but don't worry too much about the white balance. Firstly because the merging tool could let you set the white balance. But another option (that I really like a lot) is to develop my RAWs with DCRAW without white balance applied on them. That optimises the noise reduction algorithm as there is no pre-scaling on any of the RGB channels. If the result of the merge is a linear TIFF the white balance can easily be set with linear curves. I explain exactly this on this DCRAW tutorial.
David Coffin, the author of DCRAW, told me it is not recommended to develop RAW files without applying the white balance; however I have found perfect results doing it afterwards. Look at this example:

Image developed without WB:




Curves used to apply camera WB (needs the image to be a linear TIFF):




Resulting white balanced image:

This is really interesting.

I've taken the liberty to post a few CS3 HDR edits on a temporary page of my website, see http://hornerbuck.com/reference.aspx.

These edits were done very quickly using CS3 and the Photomatix tone map plug-in.

I don't profess a detailed knowledge of raw development, but isn't Tim Farrar's method the same or similar to what you propose (with a linear TIF output).
http://www.farrarfocus.com/ffdd/bracket.htm

Mike

This sounds intriguing. While I, too, don't see much difference in the results between this, digital blending and good HDR shots, the fact that it could be automated to a large degree would be a massive improvement over current techniques.

I do a lot of low-light photography (cityscapes mainly) and it's a major pain to get noise/fringe/halo-free images with digital blending using elaborate, blurred and manually adjusted contrast masks and blending modes. I dabbled with HDR but it's just too finicky and doesn't give me reproducible results, at least with my - admittedly limited - HDR skills.

Having a somewhat automatic but customizable program to produce expanded dynamic range images - for lack of a better term - would be heaven-sent. Please keep us posted on any updates!

The kind of problems he is talking about probably affect mainly saturated colors of certain hues. The RGB response of the camera is different than the RGB used in display files and mediums. RAW converters with optimized color correction need to shift hues and vary saturation, based on hue and saturation of the white-balanced image. If you don't WB before doing the full conversion, the wrong hues will be shifted and saturation-adjusted after white-balancing. Some demosaicing algorithms also work with separate luminance and chroma, and these separate differently before and after white balance.

Better terms are definitely needed. What most people call "HDR" is reall low, compressed DR display of a high DR scene. Would you call AM Radio "HDR"? No, but it is quite analogous to what is called "HDR" in digital photography.

A simple linear image with high dynamic range is really only an image with low noise in the shadows; an image in which tones are usable many stops below the maximum signal level.

yeah! It is exactly the same concept, I am glad to see someone tried it on PS. In fact in my website I pointed this:

"Another method much faster [compared to manual edition] and precise, although I ignore if possible in PS, would be to set in an upper layer the overexposed image and select on it all those individual pixels whose level in any of three channels is beyond some threshold. Done this, these pixels that would represent those burn pixels on any channel would be deleted letting see the corresponding pixels in the lower layer which would come from the correctly exposed shot"

"Otro método mucho más rápido y preciso, aunque ignoro si es realmente posible en PS, sería colocar en la capa superior la imagen sobreexpuesta corregida y seleccionar en ella todos aquellos pixels cuyo nivel en alguno de los tres canales supere cierto umbral. Hecho esto estos píxels que representarían aquellos que quemamos en algún canal, serían borrados dejando ver los píxels equivalentes de la capa inferior que corresponderían a la toma original correctamente expuesta."

I simply automated the trick so the user just needs to select the RAW files and click a button.

I see, so it's a matter of correct tones achieved. Very interesting John.




COMPLETELY agree. The problem is that 99% of people think of HDR (which actually just means 'high dynamic range') as tone mapping. That's why I prefer not to enter this battefield and not use the HDR term.

Well done gordon, it's the same result. I must confess I never managed to get realistic images from Photomatix that's why I gave up. Surely the Photomatix code did the same as my routine, whenever it was possible it took noise-free pixels from the overexposed image for the ouptput.

Very interesting approach ... useful on images that technically fall within the dynamic range of the sensor, but leave noise in the shadow (ie, not really HDR).

Using HDR to accomplish sounds good, but a program that actually works at the pixel level on the raw file would be a great utility.

The links in this thread have helped me understand HDR a little bit better as well. Thanks.

Wouldn't it be wonderful if they could incorporate the technique right into the camera. i.e.: with one click of the shutter have the camera take two exposures, automatically adjusting the sensitivity for each, and without having the mirror move twice.

Two points about this. First the new Silverfast does exactly this by using two scans of different exposure. BUT Sub $1000 flatbed scanners have huge stability problems and great difficulty aligning images. Because of heat expansion and cheap step motors the prosumer flatbeds have difficulty making two scans exactly the same length and there is some loss of resolution though noise is virtually iliminated.

Kirk,

I was interested in using the Silverfast S/W you refer to with my Minolta Multi Pro (non flatbed) film scanner (6x7 film) until reading about sharpness degradation due to mis-registration. Here is a reference:

http://tech.groups.yahoo.com/group/multipro/message/2991

I know from experience that the Multi Pro hardware is capable of producing two very similar scans that can be very accurately aligned in Photoshop, so perhaps it is (or was) a Silverfast problem that prevents them from being aligned at scan time ?

Do you know if this problem has been fixed in a newer version of Silverfast?


Ken

2 more samples. The interest of this case is to check how the algorithm performs in presence of moving scenes (water flowing). As most of the pixels of the river were taken from the overexposed image, the result was not affected.

Real LOG histogram showing this scene was 7.5 f-stops wide. Really a hard shot as the sun was about to set so contrast was severe.




Original:




Result:

Good point. With a bit more effort, though, the two exposures could be interleaved line-by-line to eliminate alignment problems. Step to first line position, acquire scan line with short exposure, acquire scan line with long exposure, step to next line position, repeat. At the end, you have two images with different exposure but perfect alignment, even with a poor stepper. I don't think anybody actually does this.

Some cameras allow exposure bracketing with mirror lock-up and a self-timer, but I'm not aware of any current cameras doing that without releasing and cocking the shutter for each exposure.

The mode you're suggesting seems to have rather few benefits over the use of multiple shutter releases.

Another useful technique is described here:

http://photoshopnews.com/2007/03/27/image-...p-cs3-extended/

With the Align Layer's command, you don't need a tripod (if you're careful) and once you use Median on the multiple SO, the noise is greatly reduced.

Hmm, that seems useful, although it seems to require quite a few exposures to achieve that usefulness.

The requirement for the extended version of CS3 is also a bit bothersome; Adobe doesn't appear to provide an upgrade from CS3 to CS3 Extended.

Actually in tests I did at the same exposure, the noise reduction was profound!

There's a back door:

http://photoshopnews.com/2007/05/14/bought...-want-extended/

I am curious if the same result can be achieved in CS2 with layer blending & image calculations applied?

A very interesting technique, especially for its simplicity. Have you compared your results from those delivered by PhotoAcute (http://www.photoacute.com)? Yes, PhotoAcute can do other things, but the noise reduction is one of the most interesting in it, and it'd be interesting to see if your method can deliver the same.

-Lars

Have you got this software to work. I've downloaded the trial and thrown batches of seven bracketed Canon 1DII raw files at it without much success - gets 40 seconds into processing and crashes every time...

Looks very interesting if they can get it to work.

As you say it does a lot of things apart from noise reduction, but regarding this topic I am almost sure it cannot get the same good results as pixel selection. Think that what I do here is to take shots with a very different exposition, so the overexposed version will have much less noise (~2^4=16 times less noise with +4EV second shot) than the other. Nothing can achieve better noise reduction than JUST taking the pixels of the overexposed version whenever possible.
What PhotoAcute does is a different concept: it does not require a difference in exposition among the different shots, it just needs a lot of shots, the more the better. Through some mean or more llikely median filter it achieves noise reduction and other advantages.