LLEMamba: Low-Light Enhancement via Relighting-Guided Mamba with Deep Unfolding Network

TL;DR

The paper tackles low-light image enhancement by integrating Retinex theory with a deep unfolding optimization framework. It introduces LLEMamba, which embeds ADMM-based iterative updates for reflectance and illumination into a deep unfolding network and uses Relighting-guided Mamba (IFBMamba) and illumination-aware bidirectional processing to enhance global context. Experiments on LOL-v1, LOL-v2, and SID show state-of-the-art PSNR/SSIM/LPIPS and reduced distortion compared with Transformer-based and other Retinex-based methods. The method offers scalable, interpretable multi-iteration enhancement suitable for real-world deployment and sets a new direction for combining Retinex unfolding with efficient Mamba backbones.

Abstract

Transformer-based low-light enhancement methods have yielded promising performance by effectively capturing long-range dependencies in a global context. However, their elevated computational demand limits the scalability of multiple iterations in deep unfolding networks, and hence they have difficulty in flexibly balancing interpretability and distortion. To address this issue, we propose a novel Low-Light Enhancement method via relighting-guided Mamba with a deep unfolding network (LLEMamba), whose theoretical interpretability and fidelity are guaranteed by Retinex optimization and Mamba deep priors, respectively. Specifically, our LLEMamba first constructs a Retinex model with deep priors, embedding the iterative optimization process based on the Alternating Direction Method of Multipliers (ADMM) within a deep unfolding network. Unlike Transformer, to assist the deep unfolding framework with multiple iterations, the proposed LLEMamba introduces a novel Mamba architecture with lower computational complexity, which not only achieves light-dependent global visual context for dark images during reflectance relight but also optimizes to obtain more stable closed-form solutions. Experiments on the benchmarks show that LLEMamba achieves superior quantitative evaluations and lower distortion visual results compared to existing state-of-the-art methods.

Figures (7)

• Figure 1: Example of (b) URetinex-Net wu2022uretinex, (c) Retinexformer cai2023retinexformer, and the proposed (d) LLEMamba for reconstruction error (i.e., color distortion, noise, artifact) comparison in the light area. In contrast to existing state-of-the-art low-light enhancement methods URetinex-Net, Retinexformer, our LLEManba not only restores the bright areas without overexposure but also relights the dark areas without noise and color distortion. This benefits from the stability of Retinex optimization and the sustainability of long-range dependencies.
• Figure 2: Overview of our proposed Relighting-guided Mamba with Deep Unfolding Network. (a) In the Deep Unfolding Network for Relight, the input is first decomposed by convolutional layers. Then decomposed images are optimized by deep unfolding layers in $k$ iterations to output the enhanced images. (b) The reflectance in $\mathbf{P}^{(k)}$ is processed by Relighting-Guided Mamba. The Relighting-Guided Mamba is a U-shaped network with Illumination-Fused Bidirectional Mamba (IFBMamba) as the backbone. (c) Then, the reflectance is processed by IFBMamba. The input is flattened and input into bidirectional Mamba after fused with Illumination.
• Figure 3: Scanning order of Mamba and Bidirectional Mamba. A bidirectional scan mechanism gathers global color context more accurately.
• Figure 4: Visual comparison on LOL-v2wei2018deep dataset. The brightness contrast between objects is well preserved in LLEMamba.
• Figure 5: Visual comparison on LOL-v2wei2018deep dataset. LLEMamba accurately captures the texture details of objects.