Illuminating the Shadows: Enhanced Low-Light Image via a Retinex-based Model with Color Equalization

Zhenbing Liu, Tao Fan, Weidong Zhang, Wenhao Wang, Jieyu Huang, Rushi Lan, Haoxiang Lu

Abstract: Lowlight images typically encounter low visibility and contrast as well as blurry details. However, most existing low-light image enhancement (LLIE) methods introduce observable color casts and unnatural-looking visibility. To address these challenges, we propose a feasible LLIE method, named RCE. Specifically, a feature hybridization attention module (FHAM) with over-lapping cross attention block and window-based self-attention strategy is developed to explore the semantic spatial relationship of pixels for generating artifact-free illumination and reflection components. We also carefully design a light enhancement module (LEM) and a denoising module (EM) to refine these two components. Meanwhile, we employ a carefully designed gate recurrent unit (GRU) and color equalization module (CEM) to balance light enhancement and color correction. Our RCE can generate comparable PSNR/SSIM/LPIPS scores for the LOL v1 and MIT-Adobe 5K datasets, whose scores are 24.74/0.8755/0.0946 and 24.91/0.9072/0.0429 respectively, benefiting from our carefully designs. Extensive experiments demonstrate that our RCE outperforms state-of-the-art methods in qualitatively and quantitatively assessments for LLIE task. Furthermore, our RCE shows the potential benefits to object detection in the lowlight condition.

External IDs:dblp:journals/eswa/LiuFZWHLL26