Researchers with the University of Illinois Urbana–Champaign and Intel have developed a deep neural network that brightens ultra-low light images without adding noise and other artifacts. The network was trained using 5,094 raw short-exposure low-light and long-exposure image pairs—the end result is a system that automatically brightens images at a much higher quality than traditional processing options.
The deep learning system was detailed in a newly published study that points out the limitations of alternative “denoising, deblurring, and enhancement techniques” on what the team calls “extreme conditions,” such as low-light images that are too dark to discern without processing.
Using traditional methods to process these images often results in high levels of noise that isn’t present when using the machine learning technique:
The team used images captured with a Fujifilm X-T2 and Sony a7S II, and also demonstrated the system on photos taken with an iPhone X and Google Pixel 2 smartphone. High-resolution comparison images are available here, and a PDF of the full study can be found here.
This is the latest example of machine learning ‘AI’ being used to automatically enhance images—ideally speeding up post-processing tasks while reducing the user’s workload. Last year, for example, a system called Deep Image Prior was demonstrated using an image’s existing elements to intelligently repair damage, and Adobe and NVIDIA are both working on AI-powered Content Aware Fill.
Articles: Digital Photography Review (dpreview.com)