AI is Taking Low-Light Photography to the Next Level

When shooting a photo in low light, a low-ISO long-exposure photo requires a stable camera and blurs movement in the frame while a high-ISO short-exposure photo can be plagued with noise and poor quality. Now AI is bridging the cap, opening the door to low-ISO image quality while shooting at faster shutter speeds.

A group of researchers at the University of Illinois Urbana-Champaign and Intel have published a new paper titled Learning to See in the Dark. It explains how they trained an AI to do low-light image processing and produce results that are much cleaner and more usable than traditional high-ISO photos.

The team put together a set of photo pairs, with each pair containing a RAW short-exposure photo and a long-exposure version.

Short exposure photos (essentially black) behind each long exposure photo used as the reference (ground truth).

The neural network was then trained with this pairs to learn how to recreate a long-exposure equivalent using a short-exposure RAW photo.

The results are remarkable: RAW photos processed with the trained AI were able to produce photos that had much less noise and much better color reproduction than photos boosted to high ISOs with the “traditional image processing pipeline.”

Example 1

A photo captured by a camera at ISO 8000.
The same scene outputted by the camera at ISO 409600.
The result of passing the ISO 8000 photo through the new AI system.

Example 2

Fuji X-T2, ISO 640, f/8, 1/30s, traditional pipeline and scaling.
The same raw data processed with the AI pipeline.

Example 3

Sony a7S II, ISO 2000, f/9, 1/10s, traditional pipeline and scaling.
The same raw data processed with the AI pipeline.

Example 4

Sony a7S II, ISO 2000, f/9 1/10s, traditional pipeline and scaling.
The same raw data processed with the AI pipeline.

Example 5

Sony a7S II, ISO-640, f/13, 1/10s, traditional pipeline and scaling.
The same raw data processed with the AI pipeline.

Example 6

Sony a7S II, ISO 6400, f/5.6, 1/25s, traditional pipeline and scaling.
The same raw data processed with the AI pipeline.

You can find more comparison examples on the research project’s website and read the paper if you’d like to learn the technical details behind the experiment.

“Experiments demonstrate promising results, with successful noise suppression and correct color transformation,” the researchers write, but they note that this is just the tip of the iceberg. The experiment lays the groundwork for much more exploration, including things like having the AI have what could be considered “Auto ISO” instead of having to provide the amplification ratio for each photo.

Source link

Leave a Reply

Your email address will not be published.