Maheshkumar H. Kolekar

Low-light images often suffer from poor contrast, reduced visibility, and detail loss, with artifacts and distortions under complex lighting further complicating enhancement. To address these challenges, we introduce DARF-Net, a Retinex-inspired framework that separates input images into illumination and reflectance components for targeted enhancement. Our method, which uses an illumination-guided multihead attention module as the generator of a generative adversarial network, improves the illumination map, while a variational autoencoder supplemented with a spatial attention module improves the reflectance map. Together, the two attention modules improve brightness, structure, and overall visual quality. Comprehensive tests on the LOL and SICE datasets show that DARF-Net outperforms the state-of-the-art techniques in terms of peak signal-to-noise ratio, structural similarity index, and learned perceptual image patch similarity metrics while retaining computational efficiency.