Wave-Mamba: Wavelet State Space Model for Ultra-High-Definition Low-Light Image Enhancement

1 INTRODUCTION

• Existing LLIE methods trained on sub-1K datasets do not transfer effectively to 4K scenarios, while UHD methods use costly downsampling that loses critical image information.The paper identifies preserving detail while augmenting global information as a central UHD restoration challenge.Source: §1 · pp. 1–2
• SSMs offer linear-complexity long-sequence modeling for global UHD information, but their unidirectional modeling is vulnerable to noise in low-light scenes.The authors frame applying SSMs effectively to noisy UHD LLIE as an open question.Source: §1 · p. 2
• Wave-Mamba uses wavelet transforms instead of traditional downsampling to preserve information and separate image content from noise.Its design rests on the observations that most image information is low-frequency and high-frequency information minimally affects LLIE results.Source: §1 · p. 2
• LFSSBlock models low-frequency illumination and texture globally, while HFEBlock uses enhanced low-frequency information to match and correct high-frequency details.The authors report that this design reduces computational costs while delivering outstanding performance.Source: §1 · p. 2

2.1 Low-light Image Enhancement

• Low-light enhancement approaches include illumination mapping, curve estimation, frequency-domain amplitude mapping, noise modeling, and diffusion refinement, while direct UHD use is constrained by computational demand.The cited examples include Zero-DCE [10], Zero-DCE++ [14], FourLLIE [5], and DiffLL [12].Source: §2.1 · pp. 2–3

2.2 UHD Image Restoration

• UHD restoration work spans bilateral learning, Fourier-domain methods, reversible downsampling, and transformers, with full-resolution consumer-GPU inference remaining challenging for LLFormer.The cited methods include UHD-SFNet, FourUHD [15], and NSEN.Source: §2.2 · p. 3

2.3 State Space Models (SSMs)

• SSMs and Mamba scale linearly with sequence length and have been applied to vision tasks, but Mamba-based restoration still faces UHD inference challenges.The paper aims to apply Mamba to UHD low-light enhancement.Source: §2.3 · p. 3

3 PRELIMINARIES: STATE SPACE MODELS

• An S4 model maps input x(t) to output y(t) through an implicit latent state h(t), with A, B, and C as state-size parameters and D as the skip connection.The input and output sequences are described as length L, and the latent state as dimension N.Source: §3 · p. 3
• The model’s continuous ODE is discretized with zeroth-order hold, using time scale Δ to convert continuous A and B into discrete counterparts.The passage presents this as the usual discretization approach for aligning the model with the input signal’s sample rate.Source: §3 · p. 3
• Mamba makes B, C, and Δ input-dependent and uses a parallel scan algorithm to process long sequences efficiently.The paper says this dynamic representation lets more pixels contribute to restoration.Source: §3 · p. 3

4.1 Observations in Wavelet Domain

• Wavelet decomposition preserves information during downsampling; low-frequency bands contain most image information, while high-frequency bands have less influence on low-light enhancement.Exchanging low-frequency components changes image content and histogram distributions more than exchanging high-frequency components.Source: §4.1 · pp. 3–4

4.2 Framework Overview

• Wave-Mamba uses a multi-scale UNet with wavelet downsampling on information-rich low-frequency features and enhanced low-frequency features to correct high-frequency features.Wavelet inverse transforms progressively upsample restored features.Source: §4.2 · p. 4

4.3 Low-Frequency State Space Block

• LFSSBlock applies layer normalization and VSSM to low-frequency features for long-range modeling, then uses GFFN to improve channel-information flow.The 2D-SSM scans features in four orientations, merges the sequences, and reshapes them to 2D.Source: §4.3 · p. 5

4.4 High-Frequency Enhance Block

• HFEBlock uses Frequency Matching Transformation to select similar low-frequency features, then applies attention and forward correction to enrich high-frequency features.Its FMTA and FCFN components handle frequency matching and correction, respectively.Source: §4.4 · p. 6

5.1 Implementation Details

• Training uses AdamW for 500k iterations, 512 × 512 random crops, augmentation, and L1 loss on two NVIDIA 3090 GPUs.The learning rate follows cosine annealing from 5×10^-4 to 1×10^-7.Source: §5.1 · p. 6

5.2 Datasets

• Evaluation covers UHD-LL, UHD-LOL4K, LOL, and LOLv2-Real, including paired UHD images and low-resolution benchmarks.UHD-LL has 2,150 paired 4K images; UHD-LOL4K has 8,099 pairs, including 2,100 test pairs; LOL has 500 pairs, including 15 test pairs.Source: §5.2 · p. 6

5.3 Comparisons with State-of-the-Art Methods

• Wave-Mamba achieves the best PSNR on all UHD datasets with few parameters and outperforms UHDFormer on perceptual metrics.On LOLv1 and LOLv2-Real, the authors also report state-of-the-art results; visual comparisons favor Wave-Mamba for color, texture, and artifact reduction.Source: §5.3 · p. 7, Table 1 · p. 7, Table 2 · p. 7

5.4 Ablation Study

• Ablations indicate that LFSSBlock and HFEBlock with FMT improve restoration; HFEBlocks add more parameters for relatively small performance gains.Increasing second-level LFSSBlocks did not consistently improve performance, so the authors selected Experiment 2 to balance performance and computational cost.Source: §5.4 · p. 8, Table 3 · p. 8, Table 4 · p. 8

• Wave-Mamba outperforms state-of-the-art methods in UHD LLIE with appealing efficiency.Source: §6 · p. 8
• Wavelet-based downsampling avoids information loss associated with high-magnification downsampling.Source: §6 · p. 8
• LFSSBlock uses SSM global modeling to recover low-frequency information.Source: §6 · p. 8
• HFEBlock uses low-frequency information to correct high-frequency details.Source: §6 · p. 8