Archive for September, 2014

September 30, 2014

SLR camera sensor high iso noise comparison

My needs in a camera sensor are relatively simple: they must not produce too grainy an image in the situations where I am constrained by needing a high shutter speed to capture action without blur.

I only compare digital SLRs (Single-Lens-Reflex) here as they have been proven to track action reliably for stills purposes whilst producing the highest available quality. I only include one camera with an electronic viewfinder and the typical sized sensor here; usually such cameras have much smaller ones in line with their compact ethos involved with doing away with the moving mirror inside them.

I discovered a site that provides a standardised test shot for different cameras at different exposures. Here I compare the results at ISO 6400. I have needed such an ISO at rare events indoors with an f2.8 aperture, but more often with a telephoto extension the equivalent of f4, for a shutter speed somewhere between 1/250-1/1000s.

Interested readers can look for themselves at the comparison images on digital photography review.

I do not consider cameras here with such a low rate of fire as to render them irrelevant for such action photography.

My results are of course estimates, based on real-life experience rather than digital sensor measurements. A number of themes recur:

1) The bigger the sensor, the more light it collects and the more detail and lower random signal noise is produced. FX or “full-frame” 35mm sensors collect about twice as much light as the 22-23mm DX/APS-C sensors, because by simple mathematics the FX sensor has about twice the area. This also places less of a premium on the optics to be sharp. APS-H is the inbetween format pursued by Canon in their sports-oriented cameras for several years.

Theoretically by sensor area, Nikon DX sensors should produce a photographic stop and a third more noise than FX, and Canon APS-C sensors almost a stop and two thirds. APS-H should make for two thirds stop more. A photographic stop means 2x per stop, as in a logarithmic scale. In practice I judged the worst camera bodies here were two stops worse than the best, or 4x the noise.

2) When sensors are relatively starved of light, the images they produced are grainy and the grain eventually overpowers the detail collected, especially in low-contrast detail. At the upper end of the ISO range, colour is lost unless the camera electronics have been programmed to compensate. In the days of film, contrast used to steepen and dynamic range used to shorten also. There were very few digital cameras that did not compensate for contrast changes. The dynamic range is the usable signal between the noise floor and the brightest highlight that can be collected by the electronics, and is inherent to the hardware.

Noise comes in two types: colour-indiscriminate luminance noise, like a grain lying over the image, and chrominance noise with random speckles or blotches of colour. It is unusually difficult to program a computer to distinguish colourless grain from the detail behind it, but it does much better with speckled colour noise. Indeed, this in practice saved the high-iso output of Canon cameras for several years.

3) The more pixels on the sensor, the more detail can be produced even after noise reduction. There was a school of thought that smaller pixels collecting less light would produce relatively more noise in proportion to the signal, but intelligently designed camera electronics have minimised the apparent noise in this.

4) A 35mm sensor shows a 50% wider field of view than a 22-23mm from the same lens, so that in practice for telephoto work at the extremes either cropping of the image must be performed, or a teleconverter used. However the appropriate teleconverter uses an extra stop of light, the stop of light that was originally a key advantage of the larger sensor. Teleconverted image quality is generally lowest at largest apertures, and in practice where a high shutter speed is required a high-pixel count sensor can compare favourably.

Naturally, I am in the hands of the methodology of the digital photography review test here. But I have seen enough of these cameras in real life to lend credence to it. Not all of the cameras below were included in the above test and I have used comparisons from elsewhere to place them.

Here is the reference image in the best quality available at ISO 50:

ticrop benchmark nikon d800 DSC_0082

I estimate the noise difference between the sensors in fractions of a stop from the benchmark of the popular action-oriented SLRs the 20D-50D series from 2004-2008. These performed on much the same level for several years until the famous D300 came on the scene in 2008, which hardly did better in terms of image quality but had so many other advantages. The 50D and the D300 also brought a new ISO 6400 “Hi” mode. I estimate the noise difference by looking up or down the ISO range when comparing cameras to see how much these look like the reference image. When one does enough of these one gets a feel for it. There are also handy pointers such as the D300-D700 theoretical difference being 1 1/3 stop. It is easy however for the naive to be led astray by the colour and contrast and hold a more positive impression than the output merits. Colour and contrast can after all be added to the picture in processing the raw files. Even professionals have been known to be led into overestimating the abilities of new cameras by the processed jpeg camera output. I also resized the larger image files down to a lower common denominator size to judge them.

It should be kept in mind that positive numbers below mean more unwanted noise, negative means less noise.

Impressions:

D2X (2004) .. + ½
1D III (2007) .. – 2/3
D300 (2008) .. 0
D700 (2008) .. -1 1/3
50D (2008) .. 0
5D II (2008) .. -1 ¼
7D (2009) .. – 2/3
1D IV (2009) .. -5/6
D3S (2009) .. -2
D7000 (2010) .. – ½
A77 I (2011) .. + ¾
D4 (2012) .. -1 5/6
5D III (2012) .. -1 ¼
D800 (2012) .. -1 5/6
D600 (2012) .. -1 2/3
D7100 (2013) .. -5/6
70D (2013) .. -2/3

Analysis:

The most impressive result unsurprisingly was from the most expensive SLR, the D3S, in the comparison, with the largest sensor to collect light. The score of minus two represents the fact that going up two stops of ISO from 6400 to 25600 presented an equivalent result to the reference 50D/D300 generation. In fact the D3S has none of the coarse white speckles of the D300 but is on a level with the 50D at 6400 after the 50D file has been tweaked for more colour and contrast. The D3S did betray significant camera pre-processing since its 12MP file was less detailed than that of the 10MP 1D Mark III. The overall impression given was that when the size of sensor was discounted, the progress of electronics has not reduced noise by more than one photographic stop. The achievement has been in increasing levels of detail at the same time.

But the large-sensor cameras need lenses twice as heavy to give the same field of view in telephoto, or teleconverters which degrade the image. There are still some remarkably high pixel-count small sensors in the list for those who like high detail. The D7100 was the best of them. At ISO 12800 the D7100 retained slightly more detail and colour than the 50D for only the tiniest amount more luminance noise.

The outsider of the group was the Sony A77, a camera with an electronic viewfinder using a largely transparent, fixed mirror that wastes some light but allows higher framerates. Despite the impressive pixel count, the image at 6400 did not impress. It even proved too difficult to apply noise reduction to it in a satisfactory way to obtain smoothing and retain any detail. It needed around half a stop more exposure compensation to reach the same exposure as the other cameras in the comparison. At the lower 3200 it was viable.

It was notable that the Canon APS-C cameras showed decidedly less detail in the image for their pixel counts than other brands. The 7D at 18MP showed less detail than the D7000 at 16MP and the 50D 15MP showed less than the D300 12MP for example. It can be speculated that the rather small sensor area was more demanding both of lenses attached, and noise processing inside the camera.

It was also found that high pixel count sensors retained their detail well when resized downwards, even with an intermediate noise reduction stage. The favourite of those who fiddle with their files. The high pixel count despite the noise helped the noise reduction programs discern detail. In fact, I would go as far as to say that where limited for reach by the focal length of a lens, the high pixel density sensor is the tool of choice for cropping.

Perhaps the lowest noise obtained from a 35mm sensor at ISO 6400 was the D4 :

Nikon D4, iso 6400

The best noise performance from a 23mm sensor at ISO 6400, the D7100 :

Nikon D7100, 6400

The noisiest from the test at ISO 6400, the A77 :

Sony A77, 6400

The result from the sensor with the lowest pixel count here, the 1D Mark III:

1D Mark III, iso 6400

The Nikon D4 at ISO 25600 ! :

Nikon D4, iso 25600

 

Advertisements