A deep adaptive framework for low-light image enhancement in adverse lighting conditions

Abstract Images captured in poor weather, weak lighting, and backlighting environments give rise to erratic visual conditions, which result in several visual abnormalities, such as poor visibility and low brightness. Direct enhancement using classic methods might face some challenges in drastically changing lighting conditions. Adaptability to various lighting conditions requires priors and handcrafted constraints, which come with higher mathematical and computational complexity. This work proposes a new adaptive image enhancement scheme (AIES) and embeds it within a degraded image enhancement network (DIE-Net), which consists of four sub-networks. The process in this framework is initiated with a multilayered image split-net that separates the reflection and illumination components. A self-regularized reflection weight removes reflection irregularities, and a transmission estimate-based illumination-upgradation weight produced by T-net increases illumination awareness. These weight functions are tailored across the reflection repair and illumination repair networks, respectively. In contrast to global methods, a new spatially adaptive weighting enhancement is proposed to extract the under-exposed regions in low-lighting and backlighting scenarios. The proposed DIE-Net performs efficiently following AIES adjustments to mitigate visual abnormalities. Experiments conducted in extremely dark and non-uniform lighting environments demonstrate the robust performance of our method.