Future DSLR CMOS Improvements

Digital cameras have been steadily improving for many years now, but recently progress has started to stagnate. One area I am particularly interested in is the low light (high ISO) performance of digital SLRs. There is enough light captured at high ISOs to generate decent images, but the main problem comes in the form of noise. While advances are still being made, the pace of change seems to be lowing. This is evidenced by the difference in capabilities between successive generations of cameras.

The Canon EOS 5D was a pretty ground-breaking camera when it was launched back in 2005. It was the first ‘affordable’ full frame DSLR, with a 12 megapixel sensor and was capable of shooting at up to 3200 ISO. In reality (as is always the case, even today), only images a stop or two below the max ISO were really usable. 3 years later a massive upgrade came in the form of the 5D mkII, improving low light shooting by a stop, allowing the same quality of shots with half the light. In 2012 the 5D mkIII added another half stop. Now, in 2016, the law of diminishing returns continues with the 5D mkIV showing virtually no improvement in low light performance over the mkIII. Below are what I consider to be the main areas of current tech where there is room for improvement and by how much.

Grain processing (1 – 3 stops)

In the decade since DSLRs like 2005’s EOS 5D, one of the main improvements to low light shooting has been courtesy of better noise reduction algorithms. Shooting a RAW image at ISO 3200 with my 5D mkII (2008) and 6D (2012) reveals that not a lot changed in 4 years. Cameras now have the ability to shoot at ridiculously high ISOs by film standards, yet they produce images with an extreme amount of grain at the higher end of their capabilities.

But where the eye sees noise, but the brain does not. Even though there may be a tenth of the information (and therefore, a tenth of the detail) in a high ISO image, you don’t perceive it as such. The brain can fill in detail, so it stands to reason that with a bit of clever processing, software should be able to emulate that.

 

Dark current and read noise (0.5 stops)

When reading the number of photons that have hit a cmos sensor, a camera produces *read noise*. This comes from extra electrons that are introduced into the signal as it is digitised. The read noise of all DSLRs is highest at their base ISO (usually ISO 100). The amount of light collected at low ISOs dwarfs the read noise, so images look clean. The read noise generated decreases as the ISO is pushed higher, but because there is less light collected, the signal-to-noise-ratio increases. This results in grainy images. Even in a camera that performs well in low light, such as the EOS 6D, the read noise doesn’t decrease above ISO 6400, meaning that at ISOs above that, the camera is simply brightening and increasing the contrast of the signal to get an image. In the past 5 years, cameras have gained on-chip dark current suppression. This could potentially be improved a little, but aside from removing a small amount of banding, there is no way to eliminate this noise with current tech.

A very in depth analysis of noise levels and sensor performance for various cameras can be found Roger N Clark’s website.

 

Layered CMOS sensors (1 stop)

Virtually all DSLR cameras (and phone cameras) have CMOS sensors. CMOS stands for Complementary Metal–Oxide–Semiconductor, and is basically a circuit printed on a silicon wafer (similar to the way processors are made). Silicon has been used as the semiconducting material in CMOS sensors for a number of reasons, the main ones being

a) it’s cheap and abundant (the second most abundant element in the earth’s crust – after oxygen – which is useful as manufacturing processes can be fairly low-yield at the beginning, which means throwing away a lot of dud silicon chips)
b) it produces good signal-to-noise ratios, allows for a high dynamic range, and has good quantum efficiency (captures a lot of light)
c) there is relatively little heat generated compared to other technologies (especially important for video or live view, where the sensor is being read every few milliseconds)
d) all microelectronics fabrication is currently based around silicon, permitting economies of scale.

CMOS sensitivity is currently about as good as it’s going to get in its current form, with over half the light landing on the sensor’s surface being captured and turned into image data. Even huge advances in efficiency here would yield well under a stop extra sensitivity.

The current way of recording a coloured image involves a bayer filter over the light gathering pixels of a sensor. This lets in a third of the amount of the light that could be used (each pixel only letting in one of three wavelengths of light, red, green, or blue). A layered sensor like the Foveon X3 records all wavelengths with each pixel, and could theoretically could allow in more light, but currently performs worse at high ISOs. One area where layered sensors have an advantage is that they reduce the appearance of noise at low ISOs due to capturing more detail, giving noise reduction software more image information to work with. This isn’t a technology that will allow more light to be gathered in low light, but rather one that will work alongside other improvements to bring higher fidelity results in light levels that cameras currently struggle with.

 

1 stop – dual ISO capture

The introduction of the 5DmkIII (and subsequent cameras) has seen Canon launch sensors that have the ability read half their pixels (every other line) at a higher ISO than the other half within the same photo. While this can result in some interlacing artefacts in extreme cases, most of the time it simply improves the dynamic range the camera is capable of recording by 1 stop. Canon has so far not implemented this in their firmware, so the only way to take advantage of it is to use Magic Lantern and change the software your camera runs, voiding the warranty. Check out this gallery of dual ISO images.

 

Nanocoatings (1 stop)

As lenses aren’t perfect, some light is reflected off the internal glass elements and bounces around inside the lens before hitting the sensor in a place it shouldn’t. This reduces contrast and potential dynamic range. Light absorbing nanocoatings are being developed that could mitigate this problem. *Vantablack* is a coating of carbon nanotubes created by Surrey Nanosystems that absorbs 99.97% of light that hits it.

Only images a stop or two below a camera’s max ISO are really usable. Objectively, I would say that currently 3200 is the highest ISO that gives good quality, noise-free images (after a little in-camera noise reduction or post processing). ISO 6400 is still perfectly usable, but is where a fair bit of detail starts to get lost, and dynamic range suffers a little (around 2 stops from its peak).

With a potential 4-5 stops increase in low light ability, ISOs of around 25/50k (good) and 100/200k (acceptable) should be possible – with 10 stops dynamic range, which I consider to be the minimum necessary – using current and emerging technologies. Not only would low light ability increase with high ISOs, but low ISO dynamic range should increase in daylight thanks to increased light gathering, signal-to-noise ratios, and increased contrast.

Graphene sensors could improve on this considerably, but they won’t be around for at least a decade, probably longer. Until then, a lot is riding on better software grain processing. This is a good thing, as advances in software will always outstrip the physical limitations of hardware eventually. There’s still reasons to be cheerful, we haven’t hit the sensitivity ceiling yet.