Low-light image enhancement based on hierarchical RQ-VAEs

Deep learning-based Low-light Image Enhancement (LLIE) methods aim to establish a mapping function between lowlight (LL) and normal-light (NL) images. To improve the performance of this ill-posed problem, some approaches have adopted VQ-VAEs to leverage normal-light prior knowledge. However, these approaches often depend on quantized features at a single scale, which may limit their effectiveness. In this study, we propose a novel LLIE method named HRQLLIE, which is based on a hierarchical structure of RQ-VAEs. Our method quantizes features at different scales using distinct codebooks, enabling the efficient extraction of multi-scale RQ-VAE features through a single training process. Additionally, we design a feature fusion module to fully integrate the advantages of multi-scale RQ-VAE features, further improving performance in LLIE. We conducted experiments on multiple LLIE datasets and achieved state-of-the-art results across most evaluation metrics, demonstrating the effectiveness of our method.