Adaptive Low Light Enhancement via Joint Global-Local Illumination Adjustment

TL;DR

This work tackles the problem of enhancing low-light images with uneven, region-dependent illumination. It presents a joint Global-Local Illumination Adjustment Network (GLIAN) that combines a Local Contrast Enhancement Network (LCEN) with a Global Illumination Guidance Network (GIGN), guided by an early stopping mechanism via the Local Discriminative Module (LDM) and a global attention-based GAEM. A three-stage training strategy couples patch-wise local refinement with global illumination cues, reinforced by a Refine Module to suppress artifacts. Across multiple datasets, GLIAN achieves state-of-the-art quantitative and qualitative results while maintaining efficiency, demonstrating strong generalization to real-world scenarios.

Abstract

Images captured under real-world low-light conditions face significant challenges due to uneven ambient lighting, making it difficult for existing end-to-end methods to enhance images with a large dynamic range to normal exposure levels. To address the above issue, we propose a novel brightness-adaptive enhancement framework designed to tackle the challenge of local exposure inconsistencies in real-world low-light images. Specifically, our proposed framework comprises two components: the Local Contrast Enhancement Network (LCEN) and the Global Illumination Guidance Network (GIGN). We introduce an early stopping mechanism in the LCEN and design a local discriminative module, which adaptively perceives the contrast of different areas in the image to control the premature termination of the enhancement process for patches with varying exposure levels. Additionally, within the GIGN, we design a global attention guidance module that effectively models global illumination by capturing long-range dependencies and contextual information within the image, which guides the local contrast enhancement network to significantly improve brightness across different regions. Finally, in order to coordinate the LCEN and GIGN, we design a novel training strategy to facilitate the training process. Experiments on multiple datasets demonstrate that our method achieves superior quantitative and qualitative results compared to state-of-the-art algorithms.

Figures (7)

• Figure 1: Histogram of brightness distribution in different regions of low-light images from natural scenes.
• Figure 2: The overall architecture of proposed enhancement framework, which consists of the Local Contrast Enhancement Network (LCEN) and Global Illumination Guidance Network (GIGN).
• Figure 3: The detailed architecture of Local Discriminative Module (LDM) and Global Attention Embedding Module (GAEM).
• Figure 4: Training strategy of the overall framework.
• Figure 5: Qualitative comparison of results on a dataset composed of three representative datasets: LOL RetinexNetLOL, MIT-Adobe FiveK MIT, and SICE SICE.