Non-uniform lighting images (NULIs) are a major challenge in image enhancement. Inspired by the human retina, in this paper a retina-like dynamic lighting-aware network (RDLAN) is designed to learn a weight map \(\:W(x,\:y)\:\) from NULI, and a dual gamma corrector with one brightener with γ = 0.4 and one dimmer with γ = 2.5 is developed to enhance NULI through guidance by the learned \(\:W(x,\:y)\) . After finding the structural consistency property, we revise the structural similarity (SSIM) to build a non-reference SSIM loss with weight-consistent loss and mean filtering loss. Our algorithm is evaluated on two fully reference datasets (LOL-eval15 and VE-LOL-L-cap), two non-reference datasets (DICM and MEF) and a real-world video dataset (SDE). The experimental results show that our algorithm outperforms some SOTAs of image enhancements, and that RDLAN is able to estimate better weight maps than other lighting-aware networks.
