Geometry-aware Depth-guided Representation Learning for Structure-preserving Low-light Image Enhancement

Jul 16, 2026

Upload by Liam

Low-light degradation reduces image visibility and weakens structural cues essential for visual representation and scene understanding. Existing low-light enhancement methods mainly restore appearance while insufficiently leveraging scene geometry to keep structural consistency. To overcome this, the paper proposes a Depth-guided Multi-scale Attention Network (DMSA-Net) that injects depth-related structural priors via reflectance–geometry interaction and a Retinex-based decomposition. Depth cues are inferred to guide multi-scale fusion, improving structural preservation and restoration results, and introducing the LOL-D dataset.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages28

Reading Time1 h 11 min

Introduction

Low-light degradation and its impact on representation

Traditional approaches and their limitations

Deep learning methods and remaining challenges

Priors for enhancement and the need for geometry-aware constraints

Geometry-aware Depth-guided Representation Learning for Structurepreserving Low-light Image Enhancement Fang Gaoa, Jiongkai Qina, Jiabao Wanga, Jingfeng Tanga, Ming Chengb, *, Hanbo Zhenga, Qingbao Huanga, Cheng Wub, * a School of Electrical Engineering, Guangxi University, Nanning, Guangxi 530004, China b School of Rail Transportation, Soochow University, Suzhou, Jiangsu 215500, China *Corresponding author | ARTICLE INF O<br>Keywords: | A B S T RA C T | | --- | --- | | | Low-light degradation reduces image visibility and weakens structural cues that are important for visual | | Low-light image enhancement | representation and scene understanding. Existing low-light image enhancement methods mainly focus on | | Geometry-aware representation learning | appearance restoration, while insufficiently exploiting scene geometry to preserve structural consistency. | | Depth-guided feature learning | To address this limitation, this paper proposes a Depth-guided Multi-scale Attention Network (DMSA-Net) | | Cross-modal attention | for geometry-aware low-light image enhancement. DMSA-Net introduces depth-related structural priors | | Structure-preserving restoration | into low-light representation learning through reflectance–geometry interaction. A Retinex-based decomposition module is first used to obtain illumination-invariant reflectance representations, from which depth cues are inferred to characterize scene structure under degraded illumination. A multi-scale depthguided fusion strategy is then embedded into a hierarchical encoder–decoder architecture, where depthaware attention adaptively integrates geometric and appearance features. Experiments on several benchmark datasets show that DMSA-Net achieves effective low-light restoration while improving structural preservation. Moreover, we construct LOL-D, a depth-augmented low-light dataset, to facilitate research on geometry-aware low-light vision. | | | | 1. Introduction Low-light image enhancement (LLIE) aims to recover reliable and structurally consistent visual representations from images captured under degraded illumination conditions. In real-world computer vision applications, visual sensors frequently operate in environments with insufficient or spatially nonuniform illumination. Due to inherent imaging limitations such as restricted dynamic range, photon starvation, and reduced signal-to-noise ratio (SNR) [1], the image formation process is severely corrupted. Consequently, low-light images often suffer from luminance attenuation, contrast suppression, color distortion, and noise amplification, which degrade both photometric fidelity and structural perception. Such degradation affects not only visual quality but also the reliability of feature representations required for image understanding. From the perspective of representation learning, adverse illumination weakens structural cues related to object boundaries, spatial layouts, and scene geometry, thereby reducing the discriminability and consistency of learned features. This limitation makes it challenging for enhancement models to preserve geometry-consistent representations during image reconstruction. As a result, downstream vision tasks, including object detection, scene understanding, and pattern recognition, may be significantly compromised when applied to degraded low-light inputs [2–5] . Therefore, developing effective computational models to mitigate illumination-induced degradation without relying on hardware modifications remains a critical research challenge. In this context, LLIE can be interpreted as a representation enhancement problem, where the objective is to recover informative and geometry-consistent features from degraded inputs, thereby supporting robust perception and downstream vision tasks. Early low-light image enhancement approaches primarily relied on histogram equalization techniques [6], [7] or models derived from Retinex theory [8-13] . These traditional methods enha

FAQs

What key problem does the paper address in low-light image enhancement?

Low-light conditions weaken structural cues needed for reliable visual representations, and existing methods mainly restore appearance without enough geometry-aware constraints.

What is DMSA-Net and how does it preserve structures?

DMSA-Net is a depth-guided multi-scale attention network that introduces depth-related structural priors through reflectance–geometry interaction, using depth cues inferred from a Retinex-based decomposition to guide hierarchical fusion.

Why does the paper construct the LOL-D dataset?

LOL-D is a depth-augmented low-light dataset built to facilitate research on geometry-aware low-light vision by providing depth information alongside low-light images.

Jul 16, 2026

Upload by Liam