Single Photons Actually Matter in Photography

As anyone with a little background science knowledge will know, light can be measured as discrete quanta called photons. These particles behave as both a wave AND a particle, (as an aside, the wave-like behaviour of light is responsible for diffraction, which softens photographs taken at small apertures). For now, I shall focus on the particle side of things.

Photons of different wavelengths from a light source bounce off various objects in a scene before eventually being captured and counted in a pixel on a camera sensor. The number of red, green, and blue photons that make it onto the corresponding coloured pixels on the sensor determine the brightness of those pixels. There are many, many trillions of these photons bouncing around, and the overwhelming majority of these never make it anywhere near the sensor of a camera that is trying to capture a scene. What is important to note, though, is that while the number of photons bouncing around is very large, it is also finite.

The number of photons collected by each of a camera’s pixels is recorded by an analogue-to-digital (A/D) converter. These are usually now 14-bits, meaning they can record up to 4096 (2¹⁴) shades of brightness per pixel (each pixel is either Red, Green, or Blue). Once interpolated, this theoretically allows for the capture of a 68 TRILLION colour image – 68,719,476,736 colours, to be precise (4096x4096x4096). This is usually output to a 24bit Jpeg that has 8-bits per channel (256 shades of brightness each for RGB, a total of 16.7 million colours).

How Many Photons are We Dealing With?

The dynamic range of most CMOS sensors is around 11 stops, or around a 2000x difference in brightness between the darkest and lightest pixels recorded. The maximum number of photons that my EOS 6D can capture for each pixel (its full well capacity) is 80,000, therefore, darkest pixels collect 800 photons at ISO 100. Each stop increase in ISO halves the number collected. By ISO 3200, only 2500 photons are being collected for the brightest pixels, and the darkest pixels measure only a SINGLE PHOTON. As a consequence, above ISO 3200 there aren’t enough photons to (accurately) use all the available steps in the camera’s tonal rage. Lifting the shadows in a photograph makes the problem worse, as the difference between a pixel hit by a single photon and one hit by two or three is substantial at high ISOs.

There are reasons for image noise other than the random nature of quantum particles. The EOS 5D mkIV has long exposure noise reduction which takes a dark frame (an exposure with the shutter closed) for the same duration as the previous exposure, and then subtracts the noise from the first image to cancel much of it out. However, this doesn’t work for the random phenomena of photons scattering on a sensor.

Software to the rescue!

So, hardware will soon reach its maximum potential (and we have seen a slowing in improvements in sensor performance for some time now), but software is where the best gains are to be had. The noise suppression algorithms that camera run can and will keep improving. I’m sure that not too long from now, deep learning techniques such as convolutional neural networks will be used to virtually eliminate image noise and restore some dynamic range. We will get image quality that far surpasses anything we have today.