Context-based low-light image enhancement via multiscale neural implicit representation

Understanding illumination in unknown scenarios while simultaneously supervising visual quality and computational efficiency is a significant challenge in low-light enhancement. Current methods either rely heavily on training data, resulting in poor generalization, or deepen the network to better understand images, which increases computational cost. To address these issues, we propose a method for enhancing low-light images based on a weight-sharing multi-scale implicit neural representation, termed Cliem. This method redefines the enhancement process in a zero-reference manner by mapping the 2D coordinates of underexposed images to their illumination components. Specifically, we first convert the RGB image into the YIQ color space and then enhance the illumination channel using a multi-scale implicit neural representation, leveraging a weight-sharing mechanism to reduce computational overhead. Additionally, we design a zero-reference loss function to boost the model’s generalization performance across various scenarios. Extensive experimental studies demonstrate that our approach outperforms current state-of-the-art methods in both subjective and objective evaluations. Furthermore, we apply the algorithm to downstream tasks in low-light scenarios, validating the practicality of Cliem.