Abstract
Low-Light Image Enhancement (LLIE) is a key task in computational photographyand imaging. The problem of enhancing images captured during night or in darkenvironments has been well-studied in the image signal processing literature.However, current deep learning-based solutions struggle with efficiency androbustness in real-world scenarios (e.g. scenes with noise, saturated pixels,bad illumination). We propose a lightweight neural network that combines imageprocessing in the frequency and spatial domains. Our method, FLOL+, is one ofthe fastest models for this task, achieving state-of-the-art results on popularreal scenes datasets such as LOL and LSRW. Moreover, we are able to process1080p images under 12ms. Code and models at https://github.com/cidautai/FLOL