Digital Camera Home > Digital Camera Reviews > Nikon Digital Cameras > Nikon D2H

Nikon D2H

Nikon introduces an 8 frame/second speed demon, with WiFi connectivity and an amazing new flash system to boot!

<<Test Results: Overview :(Previous) | (Next): Test Results: Color>>

Page 13:Test Results: Noise

Review First Posted: 12/18/2003

Test Results: Luminance Noise
(More than anyone probably wanted to know about image noise in the D2H... ;-)

Thanks to some prototype D2H evaluation units that apparently had improperly set ISO sensitivities, early results seemed to show that the D2H had much higher image noise than expected. (Credit is due to Rob Galbraith of www.robgalbraith.com for discovering and reporting this ISO error.) A good part of the surprise over the noise levels was due to Nikon's claims of exceptionally low noise for the LBCAST sensor, leading all of us who had hands-on with the early units to expect something fairly remarkable from the camera. Since then, the production units (on which this review is based) have had their ISO values adjusted to accurately match standards, and Nikon has subsequently refined their claims for low noise from the LBCAST technology to refer to only long exposure times. (The primary noise-related benefit of the LBCAST sensor design is in the form of significantly reduced "dark current.")

As noted earlier though, it does appear that, apart from the effect of the poor ISO calibration of the preproduction cameras, the images I posted earlier that were shot with one were not that far off the mark relative to noise levels we now see in full production models. Given the level of interest and dismay over the high noise levels of the early units, I took particular pains to examine the noise performance of the D2H relative to it's only real competitor in the marketplace, the Canon EOS-1D.

In the following charts, I was careful to compare shots from the two cameras captured under identical conditions, with exposures matched as closely as possible, subject only to variations in the two cameras tone curves. Given that the biggest benefit of Nikon's LBCAST sensor technology should be in low-light shooting, I compared noise levels at very low light levels, as well as at much brighter ones.

A word about the charts below: There are two ways you can quantify luminance noise in images, either as absolute variations in brightness units (usually assuming a scale stretching from 0 to 255), or as the magnitude of those variations relative to the tonal value of the subject. The latter approach recognizes that a variation of five brightness units will be much more obvious when the subject's average brightness is only 60 units than it would be were the average brightness 250 units. Both methods generally use the standard deviation of the luminance signal across a uniformly illuminated smooth color patch as the fundamental measurement of "noise." I've chosen the latter approach in most of my charts and graphs here, expressing noise as a percentage of the average luminance of the target swatch being measured. (In some later charts that plot noise levels vs luminance, I've shown noise levels in absolute brightness units, because I felt that the resulting graphs showed the relative performance of the two cameras more clearly, given that the x-axis was itself using brightness units as the metric.)

Test Conditions

As noted, I shot under two light conditions, the first being a relatively high level of illumination, producing exposure times of roughly 1/100 second at f/8 and ISO 200. This corresponds to an EV level of roughly 10.6 at ISO 100. (EV is a relative expression of light levels, tied to the ISO being used. While relative, it has become a common way of specifying absolute light levels, when referenced to ISO 100. Hence my usage above, even though none of the images in this test were actually shot at ISO 100.) This is nowhere near the brightness of direct sun, which is typically in the range of EV 15-16, depending on time of year and weather conditions, but is quite a bit brighter than the typical indoor sports venue. The second shooting condition was at a very low light level, of 1/16 foot-candle, requiring an exposure of roughly 20 seconds at f/2.8 and ISO 200. While I'm sure there'll be plenty of times when a photojournalist will need to shoot under conditions even darker than this, I think the 1/16 foot-candle condition will prove to be darker than 99% of the conditions most D2H shooters will need to be concerned with.

All of these shots were captured in the studio, at an ambient temperature of about 70-72 degrees F, and the cameras had been turned off for quite a while before shooting began. (So internal temperatures thus should have been quite close to the room ambient.) Shots taken at higher temperatures will show higher noise levels, while those taken at lower temperatures will show lower noise levels.

Single Shot vs High Speed

Richard LoPinto of Nikon alerted me to the possibility of variations in noise performance as function of shooting speed, which given my background in electrical engineering certainly made sense: With the camera firing continuously at 8fps, it would certainly be reasonable to expect more electrical noise in the system, what with the shutter and mirror solenoids blasting away, and the system as a whole cranking along at maximum throughput. Accordingly, in addition to noise measurements of images captured one at a time, I also looked at noise levels in images captured at full speed. To allow time for maximum electronic chaos to develop, I selected the fourth image in a rapid-fire series, shot under the "daylight" lighting condition. (Obviously, rapid-fire shooting isn't an option under low-light conditions.) See below for the results, but I was surprised to find very little difference relative to the single-shot numbers, and in some cases, even found lower image noise in the high-speed tests.

Noise as a function of subject color

I've found in the past that some cameras would show reasonably good noise performance with neutral tones (that is, shades of gray), but had more or less difficulty with particular portions of the color spectrum. (The most notable example was the original Sigma SD9, which showed fairly low noise levels on the neutral swatches of the MacBeth (tm) target I use for these tests, but which exhibited very high noise levels in reds.)

Accordingly, I decided to begin with a look at how the noise varied as a function of subject color with the two cameras in question. The chart below shows the luminance noise I measured in the primary color swatches of the MacBeth target, as well as in the "middle" gray patch, adjacent to the yellow block.

I should point out that, while the data here is interesting and does reveal some general tendencies, it doesn't cleanly isolate the variation of noise with color. - The problem is that the overall reflectance of the different color swatches varies quite a bit, with the red and blue swatches in particular being a good bit darker than the others. Given that the noise levels of the two cameras is a fairly strong function of luminance (darker swatches showing higher noise, even in neutrals), the peaks seen below in the red and blue hues may be less real than apparent. The chart isn't entirely useless though, as it does still serve to highlight the differences between the two cameras, since both of them respond more or less similarly to the variations in subject brightness between the various color swatches. In particular, the EOS-1D actually rendered the red swatch quite a bit darker than does the D2H, yet still managed to deliver a lower noise level, both in absolute and relative terms. (In fact, in all cases except the cyan and blue swatches, the 1D is dealing with a lower overall signal level than the D2H, so the results below should generally favor the D2H slightly.)

The overall result is that the 1D shows slightly lower noise overall, and significantly lower noise in some colors. The key to take away from this particular test though, is that the two cameras perform more or less similarly, and that the differences between them more or less reflect their performance on neutral tones. - From this, we can conclude that it's reasonable to focus primarily on noise performance in neutral tones for our subsequent tests, which is what we'll do.



The most obvious basis of noise evaluation (and the only one that's usually discussed), is how cameras' noise characteristics vary as a function of ISO, as determined by noise levels in neutral grays. As noted, given that the noise levels of the two cameras we're considering here more or less track each other across the spectrum, it seems valid to concentrate primarily on neutral tones in evaluating their noise performance as a function of ISO rating.

The chart below shows image noise levels in neutral grays as a function of ISO, for test images shot under the "daylight" conditions described above. To avoid problems due to the inherent variation in noise level as a function of subject luminance, the exposures for the two cameras were adjusted to produce luminance levels for the middle gray swatch of the MacBeth chart that were as constant as possible across all samples. I couldn't get them exactly identical, given that the exposure levels of the two cameras could only be adjusted in steps of 1/3 EV, but the luminance levels of the specific gray swatch used are within plus or minus a few percent of each other, close enough that the second-order dependence of noise on luminance can be ignored for our purposes.

It's interesting to note here that the D2H starts out at ISO 200 with somewhat higher noise than the EOS-1D, but the 1D's noise levels climb more rapidly with increasing ISO. The two curves cross each other at ISO 800, with the D2H delivering significantly better noise performance at ISO 3200. (Oddly, the D2H's image noise at ISO 400 shows here as being slightly lower than it is at ISO 200. This could be a statistical fluke, as I didn't have the time for the laborious process that would be required to generate error bars for these charts. Or, it could be a consequence of how Nikon changes their image processing with increasing ISO. - See my discussion later regarding color saturation vs ISO level.)




As noted earlier, Nikon themselves pointed out to me that image noise performance could well be different when a camera was firing at full speed than it was in single-shot mode. Accordingly, the chart below shows the same noise-vs-ISO results as above, only this time measured from images captured by both cameras running at their full 8 frames/second maximum frame rate. As previously discussed, I chose the fourth image from each full-speed series to analyze, to give a result that would be truly representative of shots captured as part of long sequences.

The behavior of the two cameras in this test was similar to that in the single-shot test, but with the odd result that the EOS-1D showed slightly less image noise at low ISOs when shooting at high speed. (The D2H also showed slightly lower image noise at ISO200, but the effect wasn't as pronounced as it was with the 1D.) I'm not sure of the reason for this, but I checked my measurements several times, and the effect does appear to be real.




The area in which Nikon's LBCAST sensor technology is supposed to make the biggest difference is in so-called "dark current", the leakage of charge into the sensor wells when no light is falling on the chip. This is by far the dominant noise effect in most cameras when shooting under low light conditions, so it's there that you'd expect to see the biggest improvement from the LBCAST sensor in the D2H.

As it turns out, this is emphatically the case, especially at lower ISO levels. At ISO 200, the chart below shows that the D2H's noise levels were only about a third of those of the EOS-1D. The two cameras again tracked each other fairly closely as the ISO was increased, but the relative difference between the two decreased very slightly as I went to higher ISO settings.


 

Image Noise: Visual Appearance

Of course, plots of the standard deviation of luminance only tell part of the story: They say nothing of what the noise looks like in the images themselves. Noise can be coarse or fine-grained, or can have a greater or lesser chroma (color) component. In general, most viewers find noise with a coarser structure or higher color content more objectionable than noise with the same level of luminance variation, but finer structure or lower color involvement.

The crops below were taken from the middle gray swatch of the MacBeth chart (the patch next to the yellow block). Images were captured under the "daylight" conditions described above. On the left is the full-color RGB image, followed by the red, green, and blue channels, shown as greyscale values.

Interesting to note here is that the D2H noise generally seems to have slightly less color to it, but its "grain pattern" is a good bit larger than that of the EOS-1D. Overall, not a clear win in terms of noise structure for either camera, it'll be up to individual photographers to decide how they feel about the "look" of the noise from each camera. I suspect though, that the coarser structure of the D2H's image noise is part of what has caused Rob Galbraith and others to object as strongly as they do to the D2H's noise levels.

 

ISO 200
 
RGB Image

Red Channel

Green Channel
Blue Channel
D2H
EOS1D

 

ISO 400
 
RGB Image

Red Channel

Green Channel
Blue Channel
D2H
EOS1D

 

ISO 800
 
RGB Image

Red Channel

Green Channel
Blue Channel
D2H
EOS1D

 

ISO 1600
 
RGB Image

Red Channel

Green Channel
Blue Channel
D2H
EOS1D

 

ISO 3200
 
RGB Image

Red Channel

Green Channel
Blue Channel
D2H
EOS1D

 

A "Real World" Noise Example

While I greatly prefer to use carefully standardized test images for performing close analysis of cameras like this, I'll be the first to say that artificial test targets tell only part of the story. I found this to be especially true when examining the noise characteristics of the D2H and EOS-1D. In particular, despite the fact that the two cameras performed very similarly in the tests above at ISO 800, in some "real-life" sample images from a side-by-side shoot at a basketball game, I consistently found that the images from the 1D looked much cleaner than those from the D2H, even though the shooting conditions were identical, and the two cameras were both being used at ISO 800, the point at which the noise levels in the tests above looked more or less identical. Some of this could be due to the somewhat coarser structure of the D2H's noise that I've already commented on, but this didn't seem to explain all of what I was seeing.

The shots and crops below show what I'm talking about. Both images were shot under the same conditions, at an ISO of 800, and with identical shutter speed and aperture settings. The only difference between the two photos was the lenses used, and the focal length they were set at, neither of which would have had any bearing on noise levels. These photos were shot by pro Gibbs Frazeur at a recent Georgia Tech basketball game here in Atlanta. (These photos are copyright Gibbs Frazeur, all rights reserved.) Like most pros I know, Gibbs routinely uses a slight underexposure in his critical digital shots to guard against lost detail in strong highlights, so the original images here are somewhat underexposed relative to what would look good on-screen. In the case of this set of sports shots, it turned out that Gibbs was actually exposing about a third of a stop higher than most of the other pros covering the game, most of whom were using 1/500 at f/2.8 and ISO 800. - All this by way of explaining why the shots below are manipulated slightly, contrary to my usual strict policy against "tweaking" images I display on this site in any way, shape, or form. To keep the comparison between the two cameras valid, I was careful to apply the identical correction in Photoshop to both of the images shown here, using the "levels" control to set the highlight point at 220. The net result was that the D2H's shot came out slightly brighter looking than that of the 1D, as the D2H appears to be about a quarter-stop more sensitive than the EOS-1D when set to a nominal ISO of 800. While it would thus be true to say that the D2H was shooting at a higher ISO than the 1D, the differential is quite slight, and certainly not enough to account for the noise difference in the images.

EOS-1D
ISO 800
1/400
f/2.8
70-200mm @78mm
D2H
ISO 800
1/400
f/2.8
80-200 @130mm

Adjusted Image
(Photoshop levels, highlight set to 220)
Original File

Adjusted Image
(Photoshop levels, highlight set to 220)
Original File

 

 

To my eye, the D2H's image noise is much more apparent here than is that of the EOS-1D. (And the D2H's color seems a little less bright as well, as shown by the rim of the basket.) This difference in noise performance was quite consistent with all the other photos shot at the event: The 1D's images were consistently smoother looking, had somewhat brighter color, and were overall quite easy to pick out in "blind" comparisons. (Blind as in that I had no prior knowledge of which images were shot with which camera.) I wouldn't consider the results from the D2H here to unacceptable, but that's going to be very much a personal decision, and I could well imagine others who are more sensitive to image noise than I am to use that terminology. (I suspect this is the sort of thing that led Rob Galbraith to term the D2H's image noise as "unacceptable.")

As noted though, this struck me as very odd, since my "laboratory" tests of image noise showed the two cameras as having nearly identical noise levels at ISO 800.

One thing that occurred to me as a possible explanation is that I'd noticed in my laboratory tests that the noise levels of both cameras got significantly worse as the luminance level decreased, all other variables being held constant. That is, there was much more image noise in darker swatches of the MacBeth chart than there were in the lighter ones. It did seem that I was seeing more of the image noise in the darker portions of the images above, so could this account for the difference? Perhaps the two cameras had the same noise level in the middle gray tone of the MacBeth chart, but the D2H suffered more in the shadows than did the 1D.

Since I had all the raw data from the MacBeth charts available, it made sense to go back and look at how the noise-vs-luminance curves behaved for the two cameras. The results were interesting, offering another view of both cameras' noise behavior with the laboratory targets, but unfortunately not really shedding much light on the real-world results above.


Noise vs Luminance

Data for the charts below was extracted from the same set of "Davebox" test shots as were used for the noise graphs above. Exposures as judged by the luminance of the middle gray patch of the MacBeth chart are thus as close to identical as I could make them, within the +/- 1/3 EV exposure step size of the cameras. Because the tone curves of the two cameras are somewhat different, the data points for both cameras do not generally line up directly above each other. This should be fine though, as the noise behavior ought to be linear enough between the data points that the graphs will give a pretty good idea of where the results for each camera will lie, at any point along the range of luminance values spanned by the available data. (As noted earlier, I chose to present the noise figures in these graphs in terms of absolute brightness units, to minimize the nonlinearity of the data between sample points.)

To provide the best view of the data at various ISO values, I made a separate chart for each of the three ISO values I looked at, with the vertical scale on each adjusted to best fit the data for that ISO level.

Without further ado, here are the plots of noise vs luminance:

These results are quite interesting. What immediately leaps out is how strongly absolute noise levels are a function of luminance for both cameras. This has important consequences for the common practice of reporting noise levels for competing cameras as part of digital camera reviews. - It's obvious from these results that reported noise values could be significantly in error if the luminance value of the swatches from which they were measured don't match exactly. Reporting single numbers intended to represent "noise level" is thus fraught with possible error. A single "noise" number also clearly tells only part of the story, because it's possible that a camera with relatively good noise performance at higher luminance values may perform much worse with darker subject matter, and cameras could vary quite a bit in this respect.

The second piece of information displayed here is a bit more subtle, namely that the noise performance of the two cameras on uniform gray swatches like this are actually quite equivalent at the two ends of the tonal scale, most of the differences between them being visible in the midtone range. This, incidentally, is exactly where myself and other reviewers reporting on noise performance tend to concentrate. Granted, this is where noise is most likely to be the most visible in the final images, but it does suggest that focusing solely on noise levels in midtone values misses a large part of the story. (Even in the face of this though, I suspect that most reviewers will continue to pay greatest attention to midtone noise values, given the sheer level of effort that's required to generate more exhaustive data of this sort. - I for one, don't have nearly the time to perform this exhaustive an analysis on every camera that I review.)

The Limitations of the Laboratory #1: Numeric vs Visual Results.

What all the experimental noise data above fails to explain though, is the quite obvious visual differences in noise performance between the EOS-1D and D2H, as shown above in the "real world" examples. In the laboratory tests, the D2H consistently shows equal or lower image noise at ISO 800, yet in the real world images, its noise is much more visible than that of the EOS-1D.

I think that some of what's going on here is an issue that I've commented on numerous times in the past: Our eyes seem to be much more sensitive to coarse-grained image noise than to fine-grained. I've sometimes found that certain cameras with higher levels of luminance noise produce images that look much better than other models with lower noise figures. The deciding factor seems to be the granularity of the noise, as the cameras with better looking images invariably show a much tighter pattern to their noise.

In the case of the D2H vs the EOS-1D, this appears to account for a great deal of what's going on. As an illustrative example, the image below is set up as a rollover, and shows color samples cropped from the central squares of each camera's image of the MacBeth target, shot at ISO 3200. Note: To head off comments that I'm focusing on a unrealistically high ISO range in this example, please be aware that I chose images captured at ISO 3200 purely to illustrate the visual-vs-numeric noise effect. The results at lower ISOs are much the same, but the differential between the numbers and the visual appearance is less obvious than it is here.

Overlaid on each color patch is a number showing the luminance noise in screen brightness units (0-255 range). As you roll your cursor back and forth over the image to switch between the samples from the D2H and 1D, there's no question that the noise in the D2H image is much more visible. Yet, in every swatch except the bright red one, the magnitude of the D2H's luminance noise is actually lower than that of the 1D. - Look at the dark purple swatch for the most obvious example of this. (This particular example also shows the reduced color saturation that the D2H produces at very high ISO levels, which I discuss in greater depth in the next section of this review.)

 



Frankly, having seen this, I'm about to conclude that the simple magnitude of luminance noise is virtually useless as a measure of image quality. - As the example above clearly shows, there's little or no correlation between noise magnitude and perceived image quality.

(And for those who might be wondering, measures of standard deviation of the individual RGB channels, or even the "a" and "b" chrominance channels of Lab color space don't seem to correlate any better with the visual results either. - Spot-checks of these noise measurements varied between the swatches in a fairly arbitrary fashion, sometimes giving the nod to the D2H, while at other times favoring the 1D, but never in a manner that matched the visual characteristics of the images very well.)

The Limitations of the Laboratory #2: Noise-suppression algorithm differences?

Another limitation of lab-based noise tests such as I've discussed above is that they're perforce performed on portions of the image having perfectly uniform color and illumination. As it turns out, this is exactly the sort of subject that camera noise-reduction algorithms love. Every digital camera on the market selling for more than a hundred dollars or so performs some level of noise-suppression image processing. (And in every case that I'm aware of, they perform at least some level of noise-suppression processing even when their "noise reduction" menu options are set to "off.")

Noise-suppression image processing is complex, and constitutes some of the most closely-guarded trade secrets of the various camera companies. It's a unique "secret sauce" that can help distinguish one camera from another. The basics are pretty consistent and well known though, and doubtless apply here. The main idea is for the image processor to look for areas of the image with no detail in them, and to work to suppress luminance and/or color variation in those areas. This makes sense, because it's exactly those areas in which you're most likely to see image noise. - Regions with a lot of high-contrast subject detail will tend to mask noise to a significant extent.

The trick comes in figuring out what areas of the subject should be considered to be "flat" and therefore subject to heavy-handed noise suppression. Generally anti-noise algorithms use some measure of "local contrast" as the determining factor, setting a threshold below which the noise suppression will take place, and above which the image will be left more or less alone. (This is a greatly simplified explanation of what would invariably be much more sophisticated algorithms.) A lot of the difference between different manufacturer's cameras has to do with how the local contrast thresholds are set. A good place to look for the action of noise-suppression algorithms is in hair detail in human subjects: You can often see just where a camera's noise suppression cuts in, based on what parts and how much of the hair looks "smooshed out" (a technical term ;-) in the photos. - My "Outdoor Portrait" test shot is a good one to look for this effect in.

This might be where we can find another explanation for the higher noise from the D2H in "real world" images than in laboratory tests. Real-world environments almost never have areas as uniformly toned and lit as the test swatches on a test target are in the studio. In the case of the shots from the basketball game shown above, there's a fair bit of tonal gradation across the area shown, although there's still relatively little that a viewer would consider "subject detail" to worry about. My guess is that the slowly varying tonal gradation was enough to convince the D2H's image processing that there was subject detail there that needed preserving, with the result that it backed off a lot on the noise-suppression, with the results shown.

Bottom Line Conclusions

From all of the above, it's clear that noise performance in digital cameras is a complex phenomena, and therefore not subject to easy characterization with one or two numbers.

I do think that we can draw some general conclusions about the D2H vs the EOS-1D as follows though:

1) Under laboratory conditions, and with bright lighting, the D2H generally equals or slightly outperforms the EOS-1D in terms of luminance noise levels in flat tints. The D2H pulls ahead somewhat at higher ISOs in midtone values, but both cameras are fairly closely matched in performance at the ends of the tone curve across the ISO range of 200-3200.

2) The numbers miss most of the story: Even though the numeric scores show the D2H handily beating the EOS-1D at very high ISO levels, the eye tells a much different story: The EOS-1D shots at high ISO look much cleaner visually.

3) Based on comparisons of multiple "real-world" image examples from both cameras, the EOS-1D carries a visible edge in its images, producing image noise that is less visually objectionable (regardless of how it might measure quantitatively), particularly in the middle-to-upper ISO range. Again, this holds true for shooting conditions involving moderate to bright lighting.

4) At very low light levels, the D2H figuratively beats the pants off the EOS-1D across the board, producing much lower noise levels both numerically and visually, regardless of ISO or subject tonality.

5) Image noise is in the eye of the beholder. Rob Galbraith labeled the D2H's image noise at middle ISO values "unacceptable." I don't know as I'd personally go that far, but Rob speaks from his extensive experience as a working sports photographer, and adds the observation that "this image quality characteristic seems to be the one giving working news and sports photographers the most difficulty." - So he apparently isn't alone in his evaluation. For my part, I felt that the D2H's images at middle ISO levels were a bit noisy, but quite usable. (By contrast though, Rob said that "higher ISO files continue to look noisy but acceptable", while I feel that it's at high ISOs that the D2H has the most difficulty.) Given this range of reactions between reviewers, all I can suggest for shooters considering the D2H is that they look at sample images shot with it and form their own conclusions.

Overall, it appears that the D2H's image noise is objectionable not so much because of its absolute magnitude, but more because it has a large, loose pattern to it. Whether this is a consequence of the LBCAST sensor technology itself, the particular data-clocking scheme that Nikon is using in the D2H, or simply Nikon's noise-suppression processing isn't clear. I'm hoping for the latter, as it would permit a fix via a firmware upgrade. If the noise pattern is a characteristic of the LBCAST technology, there wouldn't be a fix for the D2H, but we could at least hold out hope for improvements in future cameras, as the technology matures. Only time will tell on these fronts. As for the D2H itself, we'll just have to see if Nikon can make any improvements with firmware tweaks.


Reader Comments! --> Visit our discussion forum for the Nikon D2H!



<<Test Results: Overview | Test Results: Color>>

Follow Imaging Resource: