Limits to ISO

I came across an interesting post on the limits to ISO sensitivity in digital sensors (linked from the online photographer). The assumptions are a sensor with 100% quantum efficiency that is purely limited by photon shot noise (not a bad assumption) and a full-frame sensor with 12 MP resolution. Basically to get similar noise characteristics as a D3s at ISO 6400, it would have ISO 96,000. That's an astonishing conclusion and would be a really fun camera to play with. Read the whole post for the fairly straightforward calculations that went into it. Current sensors are nowhere near 100% quantum efficiency however, in large part due to the Bayer array layout which throws away a lot of light as well as the high reflectivity of the sensor surface. Both can in principle be solved and there are already concepts out there that do so.

High ISO is bad right?

One of the mantras you hear over and over is that to avoid noise, you should shoot as as low an ISO as possible because "high ISOs cause noise". However, this is based on a misunderstanding of how the physics of these sensors work. It turns out that actually, for equal exposure (i.e. shutter and aperture), you want to use as high an ISO as you can. The noise will be almost identical but you will be more evenly using the analog to digital converter. I know this sounds heretic, but it is very easy to prove. I took a set of shots all at the exact same exposure but with changing ISOs from 200 to 1600. I converted them in Lightroom 2.4 using default settings, but with the blacks set to zero and the noise reduction and sharpening turned off. I had to add exposure compensation to the low ISO shots in order to bring them up to the same level as the ISO 1600 one. Mouseover to see the other image (first time this might take a few seconds as the alternative image will have to be loaded).



D300, 50 mm/f 1.8, 1/5 s at f 3.2.

So can you guess which one is which? The slight color difference is coming from the hue shift inherent in the Adobe profiles (Adobe standard in this case) and are not real. You can get rid of those by untwisting them. But the noise is basically identical and the detail in the image is identical. The answer is that the mouseover image is ISO 200.

Of course, if you correctly expose the ISO 200 image, you will get far lower noise. Right now, it was 4 stops underexposed. This has nothing to do with the ISO, but everything to do with the physics. Correctly exposing at ISO 200 means exposing 4 stops longer. This means 2⁴ more light hits the sensor. Since incoherent photons follow Poisson statistics, this means that the noise increases by a factor 2² (=sqrt(2⁴)), so the signal to noise ratio increases by a factor of 2⁴/2²=4. What matters in modern digital cameras that have little read noise and thermal noise (as long as you're not doing very long exposures) is simply the amount of light hitting the sensor. To get lowest noise, you should expose longer or increase the aperture. At the same time, to avoid blowing out the highlights, this means lowering ISO. Some folks make the conclusion from this that this means that lower ISO means lower noise or the inverse that high ISO means higher noise. However, as I showed above, this is simply not true. For a single camera, it is simply the increased exposure that you need at lower ISO, not something inherent to low ISO itself. So if you have a fixed aperture and fixed shutter, it is probably better to increase the ISO until you get correct exposure, which in digital often means avoiding blown highlights that you care about, than it is to underexpose at lower ISO.

ETTR?

Conversely, if you only care about noise and have ample time, you should choose your lowest native ISO and increase your exposure until you start blowing out highlights that you care about. This is the ETTR mantra that you often hear (but that people often get the physics off wrong). See here for the canonical article that is right about the outcome, but wrong about the reason for the observation. It has nothing to do with digital binning steps, but everything with photon shot noise as I explain above. Note however, that if you are in a large dynamic range situation such as a setting sun on a landscape, you are far better off bracketing a few exposures and combining them afterwards or using a graduated ND filter than using the ETTR technique. In such a setting, ETTR usually means increased noise. This is because you have to expose to not ruin the very important highlights (e.g. the mountain illuminated by the setting sun), which means underexposing the darker foreground considerably. This induces lots of shadow noise. Using a graduated ND, or exposing separately for the highlights and the foreground solves this problem.

EDIT: I learned a lot by reading an article about this a while ago. So I spend a few minutes trying to tease it out of Google again. Here is a link to the pertinent part. The author gives the correct explanation behind ETTR and reaches the same conclusion as I reached above. To quote this excellent article:

Bottom line: Read noise at high ISO is much smaller than read noise at low ISO, in terms of the error in photon counting that it represents. Thus, better image quality is obtained for using the highest ISO for which the signal is not clipped.

Lightroom noise reduction revisited

A while ago, I wrote a post about noise reduction in Lightroom and how it compares to in-camera jpegs or capture NX (for Nikons) at very high ISOs. At that time, I concluded that Lightroom's noise reduction was weak. Over time however, I have come to appreciate the flipside of this coin. Both jpegs and capture NX conversions lose enormous amounts of detail because of the noise reduction applied. So much so that it can severely impact your image and make it look like it came from a point and shoot digital camera (i.e. terrible). To illustrate, here are some examples of this effect. I am employing the same rollover trick again as I have used before. Just roll your mouse over the titles and your browser will fetch that specific version. Be patient, depending on your connection, the first time this can take a few seconds to appear. After that, it will be instantaneous.


Capture NX default conversion
Lightroom default conversion
Capture NX with no noise reduction


And here is a different area.

Capture NX default conversion
Lightroom default conversion
Capture NX with no noise reduction


In my opinion, in both these examples, the default capture NX conversion pastes over lots of actual detail. I think the LR method actually gives a nice middle point between the images with noise reduction disabled in CNX and the ones with the default noise reduction. In the house image for example, the noise reduction completely gets rid of the structure in the fence, which is still visible in the Lightroom conversion. I do not like the appearance of the noise in Lightroom too much though. It seems to have a pasty character. Of course you should avoid looking this close too much at these superhigh ISO images. Of course a dedicated noise-reduction program will probably do a far better job than either. You can also see that although I used a camera-matching profile for the Lightroom conversions to give the same color as CNX, the yellow highlights on the house are not rendered as well in Lightroom as in CNX.

New tool for pixelpeepers

There is an interesting new website called dxomark that has a very comprehensive database of measurements on the performance of different digital cameras. This allow you to do quantitative comparisons between cameras from different manufacturers. Of course, you have to take these things with a grain of salt as invariably you'll find that the most expensive camera is by far the best in these measurements considering noise at high ISO, dynamic range and such. However, if you go out in the field and shoot an image at ISO 100 of a landscape with the camera you would not see any difference to your much cheaper one. And of course, if in such a situation you would shoot using a 4x5 camera with ISO 50 velvia or so, you would get far higher quality images than even the most expensive Canon 1Ds MkIII or Nikon D3. That said, I was curious for a while about what the actual IQ difference is between my D300 and the very yummy D700. Well, here it is. One example is the signal to noise ratio:

which is a full 4 dB better for the D700 at 200 nominal ISO, which corresponds to over a factor of 2 more noise in the D300. This certainly squares with my observation of noise at low ISOs in the D300. If on the DxOmark website, you roll your mouse over the scalebar on the right, you will see what this means for the image. Giving a very similar image to my examples.
Here is the dynamic range:

As you can see, apart from the base ISO, the dynamic range of the D700 is more than a stop better at every setting. That is impressive!

Now if any of this matters to you enough to spend thousands of bucks extra depends on your priorities. I wish I had the financial means to get a D700 and full-frame lenses but right now it is just lusting. I wish it had better resolution though, as from everything I have seen, the D700 gives exactly the same sharpness as a D300. The full-frame advantage doesn't hold up at all in that parameter. For that, use a 5D mkII, or a film camera I guess.

14-bit capture; does it do anything?

There is lots of confusion on the web about the benefits of 14-bit RAW capture. on the one hand people test it to hardly make any difference shooting test charts, on the other hand it is claimed that far more detail can be extracted from a 14-bit capture. One of the issues is that most of these tests use test charts or highly contrived circumstances, never a real situation. Therefore I decided to give you a real image. This is our garage in a very sunny day (it was Flag day when I shot this as you can see). I used default conversion in Lightroom of this 14-bit shot. In this sort of lighting, in order to not blow out the flag, you end up underexposing the garage. If you're smart in such a situation you might use strobes to light the inside of the garage. Another person would increase the "shadow fill" in Lightroom. Without those tricks, you cannot see the plethora of bikes and the fact that I really need to clean it out, which is perhaps a good thing.



The often repeated "truth" is that supposedly, with 14-bit capture, you get better detail in the shadows as you have the extra bits to play with. To test this, here are two details from the above guide image pushed 4 stops in Lightroom at 100%. I had to set the blacks to zero, otherwise the area stayed simply black. Guess which one is the 14 bits one!




Well? I'll tell you. The first one was shot at 14-bit. Surpsingly for an image pushed a full 4 stops, apart from being slightly brighter (even though the shutter and aperture are the same 1/500 at f5.6 and ISO 100) they are identical. You might even say that the 14-bit one has slightly less detail, but that is likely due to the slightly brighter appearance which makes the contrast and therefore apparent sharpness lower. To check whether this was due to maybe Lightroom not using the full 14-bits, I tried the same thing in Capture NX with the same result. Clearly, in real situations, using a Nikon D300, it is very doubtful that you will see any benefit from shooting at 14 bits. Not having tested other cameras with 14 bit capture such as the Canon 40D, I cannot tell you with certainty that this translates, but I would very highly doubt that you could see a significant benefit there. The reason of course is that the noise in the shadows is already large enough to be in the bottom bits in a 12-bit capture, so the only thing that happens is that you're imaging the noise with higher precision - not very useful and probably a waste of card space.

Noise in bokeh areas on the Nikon D300

I usually shoot my Nikon D300 at ISO 200 (the default). However, in certain circumstances, I have noticed noise in darker out-of-focus areas (i.e. in the bokeh). You can completely remove it in Lightroom by just dialing in a luminance noise reduction of 5. However, this prompted me to check out the noise as a function of ISO in a realistic image using default RAW conversion in Lightroom instead of some silly test chart. The NEF was shot at 14-bits and using only lossless compression. I used a 50mm f1.8 lens at f2.8. Here is the awesome scene I shot:



Here is a 100% crop at ISO 100


ISO 200:


Clearly there is far more noise in the bokeh. There is very little chance you will ever see that in a print and as I said, you can take care of it easily with just a tiny bit of luminance noise reduction, but it might be good to be aware of it.

Now for completeness:

ISO 400:


ISO 800:


ISO 1600:


This is where I reached the maximum shutter speed on the D300 (1/8000 s), so I cannot show you any more. The D300 does very well even at ISO 6400 as I reported before. At the low ISOs, Capture really doesn't do much better than Lightroom in noise reduction for those who are curious to know. However at high ISOs (>800), Capture does far better.

Interestingly, if you look at the filesizes of the jpeg exports I made from the 100% crops at high quality, you can see the exact same trend. Jpeg size, due to the way the compression operates, correlates with noise as the agorithm cannot distinguish between noise and real information. Here is the list:

ISO Size (KB)
100 148
200 160
400 172
800 188
1600 212

Since the scene and lighting was exactly the same in each image, the only thing that is different is the amount of noise in the image. Amazing that the effect is that clear.

Next: Analysis of the filesize in terms of noise

High ISO noise in Lightroom and Capture NX

Together with the new D300 came a free copy of Capture NX, Nikon's dedicated RAW converter software. Apparently Nikon wants to push the use of this program. First, after trying it a little let me say that this program still has by far the worst user interface of anything I have ever encountered. In its current form it is a major disaster. The side panels are horribly designed and the way everything works is designed as an 80's X-windows program. Horrifying to say the least. What Nikon does know very well is how to deal with the data coming from their cameras. By default it gives great color and contrast. One other thing it does well is high-ISO noise. I shot the following image at ISO 6400 with the D300 and loaded it as a .NEF in both Lightroom (2.0 beta) and Capture NX. This is the default conversion (the camera was set to Vivid picture style and Capture picks that up, Vivid adds some saturation some sharpening and some contrast)



You see this camera is usable at this ISO. There is a little noise in the dark blue sky but pretty nice. Amazing how far this has come, the image is better than film at ISO 800. Now for some 100% comparisons. This is Capture default conversion



Pretty amazing if you ask me. The noise is very grain like as if it is film.
Now Lightroom. Left, default Lightroom conversion, right optimized noise reduction (about 50/50 in both luminance and chroma)


As you see, Lightroom's conversion is far noisier by default and it is typical digital noise. The more optimized conversion is still noisy and becomes splotchy. Not very good. You also see strange speckling around the lights in both images that is not present in the Capture conversion.

Conclusion: Not surprising since the software comes from the camera maker, Capture NX deals very well with high ISO noise. If only the program were not so horrible in its use that I dread opening it every time, I would use it all the time and import tiffs or jpegs into Lightroom from it.

EDIT: because of a request below, I added the Mac OS X (i.e. identical to Aperture - my trial license for Ap 2 expired) default system RAW library conversion. It looks like this:


Very similar to Lightroom default version but not as good as Capture NX's. Somehow the noise seems to be a little more even than in the Lightroom conversion, so perhaps it would be easier to deal with, but I digress.