Unsupervised Low‐Light Image Enhancement via Adaptive Dual‐Path Frequency Network

ABSTRACT Real‐world low‐light images face multiple challenges. Non‐uniform illumination, under‐exposed regions and severe texture degradation are common. These factors make reliable enhancement difficult to achieve. Many existing unsupervised methods rely on simplified illumination assumptions. Others use coarse pixel‐level constraints. Under spatially varying illumination, these approaches often lead to over‐enhancement. They can also cause a loss of structural details. This paper introduces DAF‐Net (dual‐path adaptive frequency network). It is a lightweight unsupervised framework for low‐light enhancement. The framework explicitly models multi‐attention feature refinement. It also decouples the estimation of enhancement parameters and intensity. The network processes features through a sequence of attention modules. These include illumination‐aware attention, adaptive frequency attention, dual‐path context attention and dynamic recovery attention. This design effectively extracts under‐exposed regions, frequency‐sensitive features, contextual dependencies and spatial details. The framework estimates enhancement parameters and spatially varying intensity maps separately. These two branches are jointly optimized. This provides fine‐grained control over colour adjustment and enhancement strength. Three refined unsupervised loss terms work together. They regulate texture preservation, illumination smoothness and exposure distribution. A perceptual consistency loss serves as an auxiliary constraint. It helps maintain the semantic coherence of the enhanced image. Experiments are conducted on LSRW, MIT‐Adobe FiveK and LOL‐v2‐real datasets. Results show that DAF‐Net delivers stable and competitive performance. It achieves average LPIPS, PSNR and SSIM values of 0.169, 19.591 and 0.707, respectively.