Noises Off: Breaking the exposure triangle paradigm with AI denoising

The ISO range of modern digital sensors has increased steadily, enabling faster shutter speeds for a given aperture. Eventually though, under dim conditions, even today’s sensors fail to gather sufficient light, leaving us with noisy images. However, just like the cavalry, a new image processing technology is riding over the hill to rescue these grainy photographs: AI denoising.

The exposure triangle

An early step on the journey to becoming a competent photographer is grasping the exposure triangle. Recognising the interdependence of aperture, shutter speed, and ISO is fundamental to achieving correct exposure. All three parameters have seen impressive advances in recent years. Thus, improvements in lenses, including aspherical and apochromatic designs, mean sharp images at wide apertures. Improvements in shutter technology, including electronic shutters, allow freezing of even high-speed motion.

But perhaps the most significant change has been in the usability of higher and higher ISO ranges. For example, the Leica Q2 Monochrom has a maximum available ISO of 100,000. Despite these impressive numbers, eventually the laws of physics and paucity of photons win out. Under very low light, the ability to capture noise-free images becomes limited.

Photographers taking nighttime shots with a camera possessing modest or no in-body image stabilisation (IBIS) often resort to fast shutter speeds to avoid camera shake. Even with wide open apertures, for example f/1.7 on the Leica Q2, and a hand-held shutter speed of 1/50s, this can mean an ISO of 3200. You face a similar scenario if using a fast shutter speed to freeze motion without the help of bright daylight. Once exposure, shadows, and highlights have been adjusted in Lightroom, the resulting image can be noisy. This can manifest as graininess and/or colour distortion.

AI denoising to the rescue

Are we stuck with the limited ‘correct exposure space’ defined by the interdependence and physics-based boundaries of each parameter in the exposure triangle? It seems not.

Those of you who use Lightroom to process your RAW images might have noticed a new feature. Several months ago, a new ‘denoise’ option appeared under the ‘Edit’ panel. As the name suggests, this offers users a means of removing noise from their images. The technology underlying this new feature is artificial intelligence (AI), in particular, machine learning.

I recently began using this feature, and in this article would like to share my perspectives on its effectiveness. I am also sharing ‘before’ and ‘after’ examples of images to which I applied it. Other image processing platforms also offer denoising capabilities, but I have no direct experience of them. If you do, feel free to share your impressions in the comments section.

Deep learning

Firstly, a few words about how these denoising algorithms work. Here is an overview from the Adobe team that built the Lightroom denoising app. And here is an overview from Nvidia, who produce the chips powering the graphics processing units (GPUs) used in the training of machine learning models.

Briefly, the deep learning models underlying these denoising apps were trained on huge numbers of image pairs, one noisy and the other not. The machine learning algorithms learned what non-noisy versions of noisy images should look like and can therefore create a new, ‘noise-free’ image. 

The algorithms do not understand the images in the way humans do. They operate at the level of pixel patterns. Nevertheless, they ‘know’ what a collection of pixels corresponding to a denoised image should look like, and can therefore create one.

Heavenly bodies

This image enhancement technology is related to methodology used in the world of astronomy. In that case, scientists used it to create clear images of celestial bodies from highly pixelated images captured as unmanned spacecraft hurtle past planets, moons, and asteroids. What was once only available to the boffins at NASA is now available on your home computer.

The denoising is conducted on your device using the installed GPU. The algorithms have been optimised for the typical GPUs installed on commonly used desktop and laptop computers. I run Lightroom on a MacBook Air with an M1 chip and 8 Gb RAM. The 8-core GPU is integrated into the system-on-a-chip (SoC) design of the processor. Denoising a full frame 85 Mb DNG from my Leica Q2 takes around 2 minutes.

The use of this new denoise feature therefore adds significant additional time to the post-processing workflow. Is it worth it?

AI denoising is worth the additional effort

In my opinion, the answer is a resounding ‘yes!’.

The results achieved using denoise are extremely impressive. Unfortunately, the compression routines used to create lower resolution images for posting in this article might conceal its true impact. To my eye, when looking at 2 Mb images, the effect is astonishing. I have included several pairs of aggressive crops in an effort to make it as clear as possible.

Is this cheating? Is it using AI in a sneaky way to fool the viewer of the photograph?

In my opinion, the answer is a resounding ‘no!’.  

These systems represent the latest technological advance in photography, just as superior lenses, superior shutters, and superior ISO range have expanded the creative space available to photographers.

I look forward to further developments in denoising technology, as deep learning algorithms become even more sophisticated and GPUs become even faster. Meanwhile, I will be cranking up my shutter speed and my ISO with abandon.

Read more from the author

---

Want to contribute an article to Macfilos? It’s easy. Just click the “Write for Us” button. We’ll help with the writing and guide you through the process.

---