FLOL: Fast Baselines for Real-World Low-Light Enhancement

TL;DR

The paper tackles efficient and robust low-light image enhancement (LLIE) for real-world scenes. It introduces FLOL+, a two-stage network that first enhances illumination in the frequency domain via Fourier Illumination Enhancement (FIE) and then denoises by fusing frequency and spatial information guided by an SNR map. FLOL+ achieves state-of-the-art results on real-world benchmarks (e.g., LOLv2-Real, LSRW, UHD-LL) while using roughly 10x fewer parameters and 7x fewer FLOPs, enabling HD images to be processed in real time (e.g., ~12 ms). By training on diverse real data and evaluating on unpaired datasets, FLOL+ demonstrates robustness to domain gaps and sensor variability, offering a practical, efficient baseline for real-world LLIE applications.

Abstract

Low-Light Image Enhancement (LLIE) is a key task in computational photography and imaging. The problem of enhancing images captured during night or in dark environments has been well-studied in the image signal processing literature. However, current deep learning-based solutions struggle with efficiency and robustness in real-world scenarios (e.g. scenes with noise, saturated pixels, bad illumination). We propose a lightweight neural network that combines image processing in the frequency and spatial domains. Our method, FLOL+, is one of the fastest models for this task, achieving state-of-the-art results on popular real scenes datasets such as LOL and LSRW. Moreover, we are able to process 1080p images under 12ms. Code and models at https://github.com/cidautai/FLOL

Figures (11)

• Figure 1: Comparison of model performance on the LOLv2-Synthetic lol_v2 dataset. The chart displays the PSNR value reached for each model depending on its number of parameters, represented here by the decimal logarithm of such number in millions. The size of the points is proportional to the FLOPs of each model. Methods placed in the left-upper corner are more efficient than those placed on the right-lower corner of the chart.
• Figure 2: The illumination information is retained in the amplitude component within the Fourier space uhdllwang2023fourllie. Thus, many LLIE methods operate in the frequency domain. Source image from jiangfast.
• Figure 3: Overview of the proposed model FLOL+. In the first stage, the input image is transformed into the Fourier frequency domain to enhance its illumination level. Next, during the denoiser stage, we correct imperfections such as artifacts and high levels of noise. Note that our model is trained end-to-end.
• Figure 4: Qualitative results on LSRW-Nikonhai2023r2rnet (first and fourth rows) and LSRW-Huaweihai2023r2rnet (second and third rows).
• Figure 5: Qualitative results on LOLv1retinex_net (first row) and LOLv2-Reallol_v2 (second row) datasets.