Noise adversely affects digital camera images in a number of ways. One of those ways is by obscuring detail, thus lowering the camera’s effective resolution. I’ll be looking at the impact of increasing ISO has on the resolution of Canon’s EOS 5D mark IV’s (and by extension the very closely related EOS R’s). While the results were in line with my own positive practical experiences, it was illuminating to see how much better this sensor is compared to the one in my old 5D mark III.
In 2016 I made my first set of tests looking at the relationship between noise and the effective resolution of a digital camera system. This time, I’m repeating those tests with my EOS 5D mark IV.
Further, due to the similarity of the sensors these results are also applicable to the EOS R. Though there are slight differences in performance for the EOS R at ISO 400 and below, the differences don’t matter in this case for two reasons.
First, my test chart doesn’t provide sufficient resolution to test to the capabilities of these sensors at those ISOs. Based on my experience with the 5D mark III, it’s not until some point above between ISO 800 and ISO 3200 where noise started to limit resolution to something my chart can measure.
Secondly, the differences are small enough that they’re within the realm of being photographically unimportant. Photons to Photos.net puts the difference at least than 0.25 EV at ISO 100 and below 0.1 EV above ISO 160. In either case, the signal to noise ratios at these low ISOs are far above where I start seeing noise limit resolution to concerning levels.
I used my Canon EF 70-200mm f/4L IS USM at ~100mm for these tests. I stopped the lens down to f/5.6 to maximize sharpness and depth of field. A pair of Viltrox LED lights at 4500K and max brightness provided the necessary lighting. Only the shutter speed and ISO were varied over the course of testing.
I processed the images in Adobe Lightroom Classic CC (2019.02). Sharpening and noise reduction were turned off, and the images were converted to black and white using Adobe’s monochrome preset.
The test target is the one I developed for testing and acclimating to new cameras. You an download the target from the inked article if you want to test your own camera. Mine target was printed at 12×18″ on A3+ paper. I used Canon’s Pro Platinum paper, though any high resolution paper would be suitable. If printed properly, the hyperbolic resolution wedges should be accurate as they were generated to be properly sized in the original vector file.
As with the last test, my printer’s limitations distort the target’s rendering of the highest resolution marks. In this case, the test chart’s utility ends at about 2800 lines per picture height. This works out to be slightly less than 12 MP.
However, being limited to 12 MP isn’t a problem for this test. Under most circumstances, 12 MP is more than enough resolution to satisfy a typical viewer at typical viewing distances. My own personal threshold for printing is around 8 MP, though that may vary depending on the image in question.
One final note, these measurements aren’t completely objective. I’m reading an interpreting the hyperbolic resolution wedge visually, not using software to calculate an MTF value. Whether this is a good approach, never mind the best, is debatable.
However, for my purposes, which is to determine what ISOs are acceptable for printing under what conditions, I feel this is an entirely acceptable approach.
At ISO 100 you can clearly see my target’s defects in printing. Based on my interpretation of the target, I judge the resolution is higher than 2800 lines/picture height (ln/ph). For a 3:2 aspect image, this translates to at least 12 MP.
While 12 MP may not sound all that great, especially since this is a 30 MP sensor, I want to underscore that this is limited by my test setup and target, not the camera. Applying a reasonably amount of sharpening to this image will show detail much finer than is seen here, however, beyond about 2800 ln/ph I’m not confident enough in the chart to make readings anyway.
Impressively, to me at least, at ISO 400 not a whole lot has been lost. I still judge this image to be better than 2800 ln/ph.
Increasing the ISO to 1600 and more noise starts to become visible in fine lines. However, the 5D4/EOS R’s sensor still holds detail at around or above the 2800 ln/ph mark.
Incidentally, changes in lighting, exposures, and photographic technique from the 5D3 test has prevented the camera motion issues that occurred in the previous test.
At ISO 3200 the noise starts to become more clear as it’s completely decimated any fine detail in the very high resolution part of the wedge. However, again, the 5D4/EOS R’s sensor is holding detail clearly around, or just below, 2800 ln/ph.
At ISO 6400 the noise is increasingly more obvious. However, I’m still impressed by the ability for the sensor to hold details. While not quite as good as the ISO 3200 frame, I still judge this in the range of 2700 ln/ph or so. This translates to an image just shy of 11 MP.
ISO 12,800 is about the limit I generally like shooting at with my 5D mark IV. At this point the noise has clearly decimated any detail at the higher resolutions on the wedge. I judge this to be around 2600-2700 ln/ph, trending towards the lower end of that range.
Assuming a value of 2600 ln/ph, this results in an image with around 10 MP worth of effective resolution.
Additionally, ISO 12,800 is also the first ISO where you can clearly start to see noise in the solid black area of the central focusing target.
At ISO 25,600 noise is really obvious in both the highlights and the solid black area. Additionally, you can really see the effect of noise on contrasting edges, as there’s clearly no longer a crisp line along the edge of the back focusing target (though this was noticeable at lower ISOs too).
I judge this image to show an effective resolution around 2500 ln/ph. That translates to around 9.4 MP. If you’ve played around with my minimum resolution calculator, 9 MP is around the threshold of 20/15 vision with the print viewed at the diagonal of the print.
While I generally will want much more resolution to make a print, this is adequate to produce a good looking image (though it will be visible noisy), and it’s above the resolution needed to make a 300 ppi 8×10 (7.2MP). That said, for a fine art print, a much lower ISO would still be preferable.
By ISO 51,200 it’s clear that we’re well into the territory where you need a picture far more than you need a good picture. That said, I still judge this around 2200 ln/ph, or about 7.2 MP.
In theory that’s right on the edge of having enough information to make a 300 ppi 8×10. However, as you can clearly see noise in both the white background and the solid black of the focusing target, the quality of that image wouldn’t be great.
If anything though, I’m quite surprised by the performance the sensor is turning in here. At this point, my 5D3 was turning in a measly 1800 ln/ph (~4.8 MP). While either would be adequate for photo-journalistic conditions, it’s quite impressive to me to see how much better the sensor in the 5D4/EOS R does over it’s direct predecessor.
Much like ISO 51,200, ISO 102,400 is there for when you need an image far more than you need a good image. Noise is clearly all over the place. Fine details have been decimated, and there’s white spots in the solid black area of the focusing target.
That said, I can reasonably make an argument for reading this image at around 2000 ln/ph or about 6 MP. While I certainly don’t find this an adequate resolution for fine art work, it’s certainly workable for journalistic endeavors. It’s also substantially better than the ~3 MP I was seeing in the 5D mark IV.
I want to reiterate that this isn’t a resolution test in the traditional sense. I’m looking at how noise, driven by increasing the ISO, obscures fine detail. Moreover, as I’ve said repeatedly, there are limitations to my test target and setup that cap the highest resolution I can read from the chart.
Moreover, there is some difficulty in reading the chart itself. The hyperbolic wedge was intended to be used for measuring lens resolution, where the results would be blurred not made discretely noisy. Noise produces a notable different visual artifact compared to blurring adjacent lines, which is harder to read.
Finally, while I expected the 5D4/EOS R’s sensor to be better than one in the 5D3, and this is something I’ve seen in my own photograph. I didn’t really have a handle on it in quite so obvious terms. I’ve made prints from images shot at ISO 12,800 on my 5D4 that I never would have made on my 5D3, but I had never taken the time to preform this specific test.