By: Thom Hogan
Readers know that I shoot primarily with medium format and Canon digital cameras these days. That leaves a large block of equipment inadequately covered; most notably, the Nikon-based digital cameras. I was a long-time Nikon user (some 30 years) but since the advent of digital have been working almost exclusively with Canon gear in 35mm format.
In order to give this site more balanced coverage, I asked noted Nikon expert and outdoor photographerThom Hoganif he would contribute some of his work to theLuminous Landscapesite. Thom has a very large and active site of his own atwww.bythom.com, where he’s reviewed all of the Nikon-based DSLRs, many of Nikon’s key lenses, and much, much, more. Thom’s response has been to permit me to republish excerpts from his very extensiveNikon DSLR Reportnewsletter (the latest issue runs to over 400MBs on the CD; seewww.bythom.com/d1report.htmfor more details and the table of contents for the past and present issues). I should note that Thom was one of the first outdoor photographers to “go digital,” and was formerly the editor of Backpacker magazine. He knows both the outdoors and digital SLRs as well as anyone.
The excerpt that I’m running here is actually a small part of a much larger article that attempts to analyze and put into perspective several key image quality attributes that differ between the Nikon-based DSLRs (D100, D1 series, Fujifilm S2 Pro, and Kodak Pro 14n). Besides this section on noise, the full article deals with color fidelity, dynamic range, and exposure latitude. While we don’t get any of Thom’s great outdoor landscapes in this article, we do get his insight into a subject that we all grapple with in digital: the signal to noise ratio issue that usually rears its ugly head in the shadows of our digital images.
Noise Profiles of Nikon DSLRs
One of the first questions I get asked about every new digital SLR is this: “how noisy is it?” As most digital photographers have learned, resolution generally isn’t the controlling factor in what we perceive as “overall image quality” when we shoot digital. Ctein spoke to this at length in a recent special publication for Photo Techniques, and I agree: one of the reasons why moderate megapixel cameras can produce “better-looking-than-film” 11×14” prints has to do with the better, more controlled tonality of the digital image. That’s assuming, of course, that the digital image isn’t peppered with random noise.
I’ve done noise tests on Nikon DSLRs many times before. Indeed, the very first thing I usually do with a new body is to get a handle on its noise profile. But until this article, I hadn’t run all my bodies side-by-side though the same exact test, using the same lens, and in the same light. (As readers of the newsletter know, I shot literally thousands of images for the full article; here I’m just going to boil everything down to one worst-case test.)
Above you’ll see what I saw in the viewfinder in my set up. There is very little light available here as I was trying to mimic the worst case situations I encounter in the field. I metered off the gray card and used that as my manual exposure setting. Custom white balance was set to the lighting condition using the in-camera facilities for Preset White Balance. All images were run through the manufacturer’s converter/editor. (This example, by the way, is from a Kodak Pro 14n JPEG converted to TIFF in Photo Desk.)
All cameras were set for no sharpening and the base ISO value. Why the base ISO? Because for most outdoor/nature work you won’t go above the base ISO (some wildlife shoots are an exception); on every digital SLR I’ve tested, long exposure noise is less obnoxious or more readily controlled than high ISO noise. Again, I was attempting to make this test relevant to the worst case conditions I encounter in my outdoor photography work: low light, base ISO.
I shot two RAW exposures: one at 4 seconds, another at the fastest shutter speed I could manage in the low light circumstances of the set up (typically between 1/3 and 1 second). I also shot JPEG samples at the 4 second shutter speed. Remember, this is very worst case scenario here. Light is very low, exposures are long, and the light is unbalanced (very little red spectrum in this case). Are you ready to have your mind blown and everything you thought you knew about the relative strengths of the Nikon-based DSLR cameras suddenly thrown into question?
I measured 500×500 patches (see footnote) in Photoshop and looked for the standard deviation of the histogram’s luminance channel. Here’s the first set of numbers I found:
NA = not applicable; DNT = did not test
Anything about those numbers strike you as strange? Well, except for the S2 Pro, this wasn’t what I was expecting. I expected to see the D100 as the “noisiest” camera, with the D1x and Pro 14n in between it and the Fujifilm S2 Pro. I also expected to see noise reduction reduce noise more significantly. I expected to see more change between short and long exposures. All those expectations were based upon shooting thousands of images with each of the cameras in question and casual observation of the noise in those images. Obviously, we have a lot to delve down into to get a handle on the noise attributes.
Fujifilm S2 Pro
Let’s get the S2 Pro out of the way first, since it is the easiest to describe and pretty much what I expected once I drilled down into the detail. Every time I’ve run tests on the S2 Pro I get the same results: the S2 Pro has relatively low noise levels that don’t fundamentally change between the RAW and JPEG formats, or even within a very wide range of shutter speeds.
You may remember the cover image from the last issue (well, you Luminous Landscape readers obviously don’t, so I small version of it is reproduced below), which was a several minute S2 Pro exposure. (Five minutes if I remember correctly.) Fundamentally, it appears that the S2 Pro’s chip is very quiet when light isn’t hitting it, and if a reasonable amount of light does get to the chip, it handles that with aplomb at virtually any shutter speed you’d normally shoot at. Moreover, I tried converting at both the 12mp and 6mp size, and didn’t find any meaningful reduction of noise at 6mp, though there was a very slight one.
The S2 Pro has very clean luminance (green) channels. While both the red and blue channels rise rapidly at the onset of noise, it’s typically the blue channel that limits the amount of underexposure you can tolerate and the overall dynamic range of the camera, as it rises faster than the red channel in virtually every lighting situation I’ve tested to date (some types of fluorescent appear to be the exception). Note that overexposure can produce slightly elevated noise though this is almost never enough to be a problem (you’ll run out of exposure latitude before you hit significant noise).
(Comment about the chart. The numbers across the bottom are stops of exposure. A standard gray card at the proper exposure is the zero point, and then I ran one stop increments from that exposure to get the other values. So in the case of the S2 Pro, shown here, noise is very well controlled through about five stops, then begins showing up in the blue and red channels as you move down into the shadows.)
In general, I’ve always been impressed with the S2 Pro’s noise handling, and this issue’s tests simply confirmed that once again. (I should also point out that turning in-camera sharpening up for JPEGs does tend to bring out more noise on S2 Pro images. Though the primary culprit causing that noise will be both color fringing and JPEG edge artifacts, overall noise does seem to go up more so on the S2 Pro with high sharpening than on, say, the D100.)
Kodak Pro 14n
The Pro 14n probably surprised you, didn’t it? At least at the base ISO value, the camera seems to fundamentally have no noise problems in the mid-range values, just like the S2 Pro. You may raise your eyebrow at the higher number for the noise-reduced Long exposure versus the Long exposure without it. So was I. We’ll revisit that in a moment when we get to the D100.
In terms of color, the Pro 14n is a bit like the S2 Pro: the green channel mimics the luminance in noise, while noise in the red channel comes up faster and the blue channel gains noise the fastest. The blue and red channels dictate the point at which noise becomes problematic. The blue channel produces objectionable noise as early as two stops under proper exposure, while the red channel usually holds up for another stop or so. But notice that the noise is already ramping up very early (-1EV) in all channels. I suspect that the Pro 14n actually should have been specified and linearity balanced at ISO 50. (Folk who’ve seen my posts on the Pro 14n on various forums also know that I tell Pro 14n users to “remove” the bottom two stops of exposure from their images. The Pro 14n will attempt to record as much as 10 stops of dynamic range in high contrast scenes, but those bottom stops generally have too much noise in them. Kodak appears to have realized this, as Photo Desk now has a feature that allows you to clip the low-end of your exposure. And when you use it, voila, the objectionable noise everyone complains about with the Pro 14n disappears.)
Overall, I was impressed with the noise tendencies of the Pro 14n on a well exposed shot. Unlike the S2 Pro, the Pro 14n’s tendencies above middle gray are relatively benign in terms of noise. It’s just not a camera that tolerates underexposure well.
The D1x probably also surprised you, as it did me. The first number in the charts that introduced this section (when there are two numbers) is for a 10.4mp conversion, the second for a 5.6mp conversion. You’ll note that in each case, the smaller conversion (and the JPEG, which uses the same size) is significantly better. I puzzled about why that was until I looked very closely at the image. At these exposure lengths, the D1x was already starting towards hot pixels, even though they didn’t really show up as obvious bright spots in the image. When I cranked up the exposure to 10 seconds, the hot pixels were obvious, and they were in the exact same place as the significant chunks of noise in the shorter images. Eek.
When I went the other way and shot in the sun with 1/60 of a second or faster, the D1x was back down in the same range as the S2 Pro and Pro 14n (2.56 versus 2.32 and 2.41). Thus, my worst case noise test really was a worst case test for the D1x, and as I thought about many of the images I’ve shot, I started to realize that I may have been producing this problem in my images without knowing it. So I went back and looked for a few images shot in the 1 to 2 second range. Visually, they didn’t seem noisy. But there was a just-below-visual-detection noise running in them (just use a really high Unsharp Mask setting and it’ll come popping out at you). So, I, too, learned something in these tests: my D1x is not as noise-free in those slow exposures as I thought it was. When I can, I’ll be trying to get my D1x exposures up to at least ½ second, and I recommend you do the same.
In terms of the individual channels, the D1x shows a pattern that seems to be typical for the Nikon brand bodies: all the channels tend to run extremely parallel to one another, with the green and luminance channels being only slightly better than the red and blue at the extremes. You’ll also note that there is a “trough” in the middle of the tested range where noise tends to be somewhat higher than the non-Nikon bodies, but still under control. In the graph below of one of my tests, that “trough” ran from +1EV to -3EV. Again, the blue channel tends to be the ultimate limiter of how much underexposure can be tolerated, but note that the luminance channel isn’t far behind.
Overall, the D1x’s noise performance is good. While the above chartlooks bad, note that even at -4EV the actual noise levels are still within the range that Neat Image can typically remove (in my rush to get this issue to “press,” I didn’t scale these charts the same, so note carefully the Std Dev values on the vertical axis—none of the D1x values are hitting 6 here, while they did with the S2 Pro and almost hit 10 with the Pro 14n).
Which brings us to the last strange result in my worst case tests, the D100. At first I couldn’t quite figure out why using noise reduction on the long exposure produced a worse result. Then I looked closely at the resolution chart at the top of the frame (therewasa reason why I included it, after all). It turns out that I had shot the D100 with anti-mirror turned OFF and in the short exposure it appears that the “bounce vibrations” were significant enough to average out some pixels in the grey card. When I turned the anti-mirror function on, suddenly order was restored and the noise reduced images were better than the ones without it.
Which brings me to an aside: the N80-based cameras most certainly have mirror slap problems at exposures between 1/15 and 1 second. With longer lenses (and especially macro focus), this is enough to affect the visual clarity of the image; you’ll get blurred edges at best. Note that I was using a solid ball head locked down onto a Kirk Ground Pod locked down on a sturdy table. The 70-180mm was mounted via a Kirk plate to the head, and everything tightened as much as possible. This is not a set up that should have stability problems, yet it does when the D100, S2 Pro, or Pro 14n are used at those suspect speeds. The mass of the body hanging off the back of the lens appears to be enough to compromise the stability, and this is consistent with what I’ve seen in the field—short of using a second brace to the body and a very sturdy tripod, you should either avoid these shutter speeds or use the D100’s anti-mirror function (the latest firmware of the Pro 14n also has a mirror prerelease function). The D1x, even though it’s a heavier body, doesn’t seem to have the same level of problem, though in extreme cases (600mm, for example), I have seen mirror slap affect images, so the same caution does apply, though the problem isn’t as serious.
At this point, I went back and examined the Pro 14n results and found the same thing: in the picture with noise reduction OFF, the resolution chart showed that the camera had suffered from considerable bounce.
In terms of channel noise, the D100 is somewhat like the D1x—in the usable exposure range the noise is slightly elevated compared to the S2 Pro and Pro 14n, but still very much under control. Again like the D1x we have a trough that leads to a very steep rise in noise in all the channels. Essentially, when the D100 gets “noisy,” it does so with gusto in all the channels.
Additional Noise Comments
Given the results I had with noise reduction invoked in the worst case tests, I re-ran the tests in brighter light and with much faster shutter speeds:
This was more like it. Noise reduction on both the D100 and Pro 14n triggered an overall lowering of the noise floor, though not a dramatic one.
So what have we learned? Not nearly enough, so let’s delve a little deeper.
Another way to look at noise is to look at it across the full recording ability of the camera. I’ll take some of the patches from the ColorChecker chart and run them through Photoshop’s Histogram function to get standard deviations this time. Here is what I measured:
Now things are getting very interesting. Indeed, I can say that several of the trends I see here absolutely show up in images I’ve taken:
• TheFujifilm S2 Prohas very low noise across the board until you get down low in its capture range (or underexpose significantly). Here it doesn’t really show up until we get to the Black patch, but I reran the test underexposed a stop and got what I expected: the Dark Gray patch was the break point where noise began rocket upwards, and the Black patch was well over 4.
Conclusion:Underexposure is a big noise risk with the S2 Pro, especially if you have to later pull up the exposure with changes in linearity. Push this camera right to the top edge of the histogram (and since you have the channel histograms, check them, too). There doesn’t seem to be much differential due to color—it’s only when you get three stops below the exposure mid-point that noise starts to rear its head, and it’ll happen in any color at that point. Even so, watch yourself in light that has little in either the blue or red spectrum, as the channel that gets underexposed here will almost certainly be the trigger for noise in your image (e.g., in incandescent light, the blue channel isn’t receiving much energy, gets underexposed, and becomes the primary noise contributor).
• ThePro 14nprobably surprised you, but look closer at its numbers. There is a clear trend for noise to increase as the exposure level goes down. In other words, the Black patch had more noise than the Dark Gray patch, which in turn had more noise than the Medium Dark Gray patch. This, too, is something I see in my pictures from this camera (and one of the reasons why I say that you need to raise the black point in Pro 14n images or otherwise clip the bottom of the dynamic range captured). Again, I ran additional tests at lower exposures, and the noise results moved up the patches as I expected. Note that the “black” patch wasn’t completely black (i.e., wasn’t at or near the expected 0,0,0 point), so even darker results could have been recorded, and they would have had more noise.
What I can’t explain are the color results. The Yellow noise value is the lowest Standard Deviation I’ve ever measured with this size sample—essentially it says that virtually every pixel measured across a huge number of pixels was no more than 1 value away from expected. Since CYM is derived from RGB and the RGB numbers are consistently higher than the CYM numbers, something interesting is happening the Kodak’s color processing.
Conclusion:Shadows are where noise lives with this camera. Again, pushing the exposure as high as you can without blowing highlights is recommended, but I’d also recommend that you not try to use the full dynamic range this camera can produce, as the lower stops will have more and visible noise if you do. The blue channel is definitely the primary noise trigger, though all the channels rise rapidly with underexposure. But since blue is the worst channel, watch yourself in incandescent or low Kelvin studio lighting!
• TheNikon D100is more fickle and difficult to get a handle on. Surprisingly, both the bright and dark end of the exposure range tends to produce noise. The mid-range has good noise properties, but both extremes show rises in noise levels. Again, I’ve seen this in images, though more so at the dark end than the bright. But this does explain some unusual problems I’ve had in a few clouds and other bright areas in some shots. (If you have the full issue from which this excerpt is taken, look at the noise in the cover image—it’s definitely there in the brighter areas—indeed, the detail that has values just below the midpoint of the histogram shows the lowest amount of noise, just as the chart above would predict.)
One thing to note: the Red patch showed the highest amount of noise, and this carried over a bit to the Yellow and Magenta patches (both of which have red components in them). The light for this shot was deficient in red spectrum compared to blue, and this seems to have triggered the primary noise build-up in this image. I’ve seen it the other way, too. I’m still trying to get a handle on what the “trigger” is, but since all channels in this camera tend to be living just under the point of noise tolerance and relatively equal, I’m at a loss to explain exactly what I’m seeing at this point.
Conclusion:Get the white balance and exposure right; don’t press right to the bright edge of the histogram unless you need to for dynamic range (though if it’s a choice between pushing to the right or left edge of the histogram, choose the bright right edge!). Underexposure produces noise, sure, but unlike the non-Nikon cameras, noise remains relatively the same across the exposure latitude you can correct for in Capture. Also: If I were shooting in a studio with the D100, I most certainly would want to use lights that are near daylight balance, not Tungsten. Unequal or difficult white balance situations seem to trigger noise on this camera.
• TheNikon D1x, not surprisingly, is a lot like the D100, though not quite as extreme. The mid-range again tends to be mostly noise-free, but the extreme edges of the exposure eventually become a problem. (The low number for White in this last chart may indicate that the exposure was slightly hot; when I went back and retested with a slightly lower exposure, the results more closely matched the D100, with a slight rise at the brightest patch.) Again, the Red patch shows more noise than the others, though the difference isn’t quite as distinct as it was on the D100.
Conclusion:Like the D100, the D1x does best in the middle of its dynamic range and equally well across the range of exposures you would typically correct for in Capture. I’m a little less leery of low red or blue spectrum with the D1x than with the D100, though, and overall, the results seem a bit more consistent.
You should note a couple of things that were said in passing. For instance, noise interacts with dynamic range (see the full article in my newsletter for more on dynamic range). If you have to bring up an underexposure, you’ll be pulling up noise as you do so. This is most obvious on the S2 Pro and Pro 14n, but the Nikon bodies also exhibit this, though to a lesser degree until you reach their “break point.” Also, channel noise can differ significantly, which affects color integrity.
You might also wonder what a “good” value is to see in these tests. After doing this and other similar tests often and having the ability to look at how that translates into real images, I don’t worry much when I see values below 2.0 on samples 100×100 or larger. I think it would take values across the board under 1.0 to get close to the point where I’d be impressed, and even then noise still would be lurking underneath; no current camera I’ve seen comes close to that. Even values significantly higher than 2.0 can often be dealt with by products such as Neat Image, though the higher the value, the more you have to worry about detail destruction and even color drifts. In short, all four cameras produce very usable images at regular shutter speeds across their usable dynamic range, and even at long shutter speeds and slight underexposure. There’s not a heck of a lot to distinguish the cameras at their base ISO values, though the S2 Pro seems the most noise free right up until you get to the darkest shadows.
You might ask why I don’t test the cameras at higher ISO values. Actually, I have run such tests on my cameras, but I find that as you press ISO up I see more individual sample variations between cameras than I see at the base ISO. I’ve measured one D1x that was better at ISO 640 than my personal body is at 500, for example. Thus, I recommend that you test your own camera and built a noise profile for it. Try to get a feel for what level of underexposure or ISO boost starts the rapid ramp upward (I’ve yet to find a camera where there isn’t a point at which the noise curve doesn’t suddenly arc upwards). Personally, I try to never shoot at anything but the base ISO value and also try to hard to avoid underexposure.
© 2003 Thom Hogan
That’s 250,000 pixels, so the Standard Deviation should be very reliable. Some other testers use 100 x 100 or even smaller blocks. What I’ve found is that smaller sample sizes tend to produce somewhat lower numbers than I achieved here, so be careful against pitting numbers you see here against those found elsewhere.
What’s a Standard Deviation? It’s a measure of, on average, how far from the Median value the samples lay. That’s why we make this measure on something that has an even, known color, such as those on a gray card or better still, the ColorChecker chart. A standard deviation of 0 would indicate perfect results—in other words, no deviation between any pixels in the block. A standard deviation of 2 on a gray value of 90, for example, would mean that, on average, the average difference from 90 would be 2 (e.g., 88 or 92). That could also mean that a few values were at 180 and most of the rest were right at 90. More usual, however, is that the recorded values form a standard bell curve around the expected value. That’s what I saw in most cases here, so a standard deviation of two tended to mean that more than half of the values fell with 2 values on either side of the expected value, and the rest were mostly close in. In a couple of cases I saw some far outlying values. For example, on the 4-second D1x exposure, I saw a few hot pixels. As you read the charts, remember that higher values are worse than lower values.
Thom Hogan is the former editor of Backpacker magazine. Before that, Thom worked in the high tech business, including working on the design of some of the earliest digital cameras. These days he writes about the outdoors and photography, and his work can be see atwww.bythom.com.