Dual Degradation-Inspired Deep Unfolding Network for Low-Light Image Enhancement

TL;DR

This work addresses the need for physically interpretable low-light image enhancement. It introduces a Dual Degradation Model that separates luminance and chrominance degradation and unfolds an alternating optimization into a multi-stage network (DASUNet) with space-specific priors realized via a luminance Transformer and a wavelet-based chrominance Transformer, coupled through a Space Aggregation Module. The method achieves state-of-the-art results on LOL and MIT-Adobe FiveK, supported by comprehensive ablations and runtime analyses, and demonstrates clear advantages in interpretability and efficiency. Overall, the dual-space, deep-unfolding framework offers robust enhancement with explicit degradation priors and promising applicability to other low-level vision tasks.

Abstract

Although low-light image enhancement has achieved great stride based on deep enhancement models, most of them mainly stress on enhancement performance via an elaborated black-box network and rarely explore the physical significance of enhancement models. Towards this issue, we propose a Dual degrAdation-inSpired deep Unfolding network, termed DASUNet, for low-light image enhancement. Specifically, we construct a dual degradation model (DDM) to explicitly simulate the deterioration mechanism of low-light images. It learns two distinct image priors via considering degradation specificity between luminance and chrominance spaces. To make the proposed scheme tractable, we design an alternating optimization solution to solve the proposed DDM. Further, the designed solution is unfolded into a specified deep network, imitating the iteration updating rules, to form DASUNet. Based on different specificity in two spaces, we design two customized Transformer block to model different priors. Additionally, a space aggregation module (SAM) is presented to boost the interaction of two degradation models. Extensive experiments on multiple popular low-light image datasets validate the effectiveness of DASUNet compared to canonical state-of-the-art low-light image enhancement methods. Our source code and pretrained model will be publicly available.

Figures (12)

• Figure 1: The comparison between existing unfolding methods and our proposed method. (a) the existing unfolding methods; (b) our proposed method.
• Figure 2: The architecture of our DASUNet. The low-light image $\mathbf{y}$ is transformed to luminance space $\mathbf{y}_{lum}$ and chrominance space $\mathbf{y}_{chrom}$. Then, they are respectively fed into luminance optimization stream (LOS) and chrominance optimization stream (COS). After $k$ stages, the enhanced image $\mathbf{x}$ is produced.
• Figure 3: The comparison of luminance and chrominance spaces. We can see that there are different distoration degrees in luminance and chrominance from quantitative and qualitative perspectives.
• Figure 4: The illustration of network structure of each phase. (a) Gradient Descent Module (GDM); (b) Prior Modelling Module (PMM); (c) Space Aggregation Module (SAM); (d) Wavelet decomposition Transformer; (e) luminance adjustment Transformer; (f) Feed-forward layer.
• Figure 5: The visual comparison on LOL dataset.