FAN: A Fourier Alignment Network for Low‐Light Image Enhancement

ABSTRACT Low‐light images are prevalent in many real‐world scenarios, such as surveillance, autonomous driving, and mobile photography. However, these images often suffer from low contrast, high noise, and blurred details due to insufficient lighting. These degradations not only compromise visual aesthetics but also significantly hinder the performance of downstream tasks such as face recognition and security monitoring. To tackle these challenges, we propose the Fourier alignment network (FAN), which decomposes illumination and noise into amplitude and phase via frequency‐domain analysis and fuses spatial and frequency domain features in a cohesive framework. A spatial‐domain enhancement network effectively boosts illumination and suppresses artifacts, while gated convolution aligns illumination by capturing pixel‐wise lighting offsets. We further introduce focal frequency loss to minimize frequency‐domain discrepancies and complement traditional spatial losses, which guides the model to recover accurate illumination while preserving fine details. Comprehensive quantitative and qualitative experiments on the LOL, LOLv2, and UHDLL benchmarks have validated the effectiveness of our approach. Notably, on the UHDLL dataset, FAN achieves a PSNR of 26.28 dB and an SSIM of 0.92, which outperforms previous state‐of‐the‐art methods.