PaperlayerView exact revision on arXiv

Source-linked AI summary

Bidirectional Cross-frequency Domain Interaction Network for UHD Low-Light Image Enhancement

Bo Fu, Hai Chi, Liyan Wang, Shilin Fu, Wei Wang, Kin-Man Lam, Jinshan Pan, Cong Wang

arXiv:2610.33402v1cs.CV

TL;DR

UHD low-light enhancement methods often do not jointly exploit spatial–frequency information and relationships between illumination and structure across frequencies. BiFreqNet applies reciprocal low-/high-frequency enhancement to pretrained VAE multi-scale features, and reports favorable performance across general and UHD LLIE benchmarks. In extremely low-light scenes, insufficient RGB color information can still make white-balance and color-temperature restoration difficult.

  • Problem

    Existing UHD LLIE methods often underuse complementary spatial–frequency information and cross-frequency relationships between illumination and structure.

    Source: §I · p. 1, A. UHD Low-Light Image Enhancement · p. 2

  • Method

    BiFreqNet applies a DFI module to pretrained VAE multi-scale features, using high-frequency cues for low-frequency restoration and enhanced low-frequency context for high-frequency refinement.

    Source: §I · p. 2

  • Results

    BiFreqNet performs favorably against state-of-the-art general and UHD LLIE methods across four public datasets.

    Source: §I · p. 2

  • Takeaways & Limitations

    The method uses reciprocal low- and high-frequency information to support illumination restoration, noise suppression, and reconstruction of structures and fine details.

    Source: §VI · p. 12

    In extremely low-light scenes, insufficient RGB color information makes accurate white-balance and color-temperature restoration difficult and can lead to color tone shift.

    Source: §V · pp. 11–12

Abstract

from arXiv · show

Existing Ultra-High-Definition (UHD) low-light image enhancement (LLIE) methods typically rely on either spatial- or frequency-domain modeling, or independently process different frequency components for efficient restoration. However, they often fail to jointly exploit complementary spatial--frequency information and the inherent relationship between illumination and structural information across frequencies, leading to inconsistent illumination, residual noise, and detail artifacts. In this paper, we propose a novel Bidirectional Cross-frequency Domain Interaction Network (BiFreqNet) for UHD LLIE. Specifically, BiFreqNet incorporates a Dual-domain Feature Interaction (DFI) module into the multi-scale representations of a pretrained variational autoencoder (VAE) to perform reciprocal enhancement between low- and high-frequency features. DFI consists of a Spatial and Spectral-Aware Low-Frequency Enhancement (S$^2$LFE) block and a Cross-Frequency Enhancement (CFE) block. S$^2$LFE exploits high-frequency structural cues to guide low-frequency restoration and jointly models spatial and spectral information to recover global exposure and spatially varying illumination while suppressing sensor noise. CFE further leverages the enhanced low-frequency contextual information to guide adaptive high-frequency refinement, enabling faithful reconstruction of structures and fine details. Extensive experiments on four general and UHD LLIE benchmarks demonstrate that BiFreqNet achieves competitive or superior performance compared with state-of-the-art methods.

I. INTRODUCTION

UHD low-light enhancement must restore illumination while preserving fine details, but existing approaches do not fully exploit complementary spatial–frequency information or dependencies between frequency components. BiFreqNet uses reciprocal cross-frequency enhancement, and experiments across four public UHD and general LLIE datasets report favorable performance against state-of-the-art methods.

I. INTRODUCTION

  • UHD LLIE must restore illumination and preserve fine-grained details, while conventional-resolution methods can incur high computational costs on UHD inputs.The supplied text identifies both restoration capability and UHD processing efficiency as challenges.Source: §I · p. 1

A. UHD Low-Light Image Enhancement

  • Existing UHD methods using one domain or separate frequency components underuse spatial–frequency complementarity and cross-frequency links between illumination and structure.The paper identifies this gap as a source of suboptimal restoration and describes low- and high-frequency restoration as interdependent.Source: §I · p. 1, A. UHD Low-Light Image Enhancement · p. 2

I. INTRODUCTION

  • High-frequency structures provide boundary and local-contrast cues for low-frequency illumination restoration, while restored low-frequency context guides detail reconstruction.BiFreqNet’s visual comparison is described as producing more natural illumination and clearer structures and details.Source: §I · pp. 1–2, Figure 1 · p. 2
  • BiFreqNet integrates a DFI module into pretrained VAE multi-scale features for reciprocal enhancement between low- and high-frequency representations.S2LFE uses high-frequency cues for low-frequency illumination restoration and noise suppression; CFE uses enhanced low-frequency context for high-frequency detail refinement.Source: §I · p. 2
  • BiFreqNet performs favorably against state-of-the-art general and UHD LLIE methods across four public datasets.The evaluated benchmarks include both UHD and general LLIE datasets.Source: §I · p. 2

B. Variational Auto-Encoders

The paper places BiFreqNet’s bidirectional interaction on multi-scale features from a pretrained VAE, then fuses enhanced and original features for reconstruction. It positions this approach against VAE-based UHD restoration that leaves low-frequency illumination and high-frequency structure interaction underexplored, and against state-space methods that do not jointly use spatial, spectral, and structural information.

B. Variational Auto-Encoders

  • Existing VAE-based UHD restoration methods mainly improve latent representations or compensate for compression loss, leaving cross-frequency interaction between illumination and structure underexplored.BiFreqNet applies bidirectional interaction to multi-scale VAE encoder features for illumination restoration and detail recovery.Source: B. Variational Auto-Encoders · p. 3

C. State Space Models

  • The paper identifies a gap in jointly modeling long-range spatial dependencies for varying illumination, spectral information, and high-frequency structural guidance.Its stated approach uses Mamba for long-range dependencies across illumination regions while incorporating spectral information and structural cues.Source: C. State Space Models · p. 3

A. Overview

  • The two-stage pipeline pretrains a VAE for compact latent representations and multi-scale features, then enhances those features through bidirectional interaction and reconstructs the image.In the enhancement stage, the frozen encoder and decoder surround DFI-based feature enhancement and feature fusion.Source: A. Overview · pp. 3–4

B. Dual-Domain Feature Interaction (DFI) Module

The DFI module combines spatial and spectral modeling of low-frequency features with bidirectional interaction between low- and high-frequency features. Its S2LFE and CFE blocks use structural guidance for low-frequency restoration and recovered illumination context for high-frequency refinement.

B. Dual-Domain Feature Interaction (DFI) Module

  • DFI stacks S2LFE blocks for low-frequency restoration, then uses CFE to refine high-frequency details with recovered low-frequency context.This combines spatial-spectral modeling with interaction in both directions between the frequency representations.Source: B. Dual-Domain Feature Interaction… · p. 4
  • At each VAE encoder scale, Haar wavelet decomposition separates a low-frequency feature from three directional high-frequency features for enhancement and reconstruction.The reconstructed enhanced feature is concatenated with its encoder feature, projected, and passed to the frozen VAE decoder.Source: B. Dual-Domain Feature Interaction… · pp. 4–5
  • S2LFE jointly models low-frequency features in spatial and spectral domains, using high-frequency structural cues to guide illumination restoration and sensor-noise suppression.The guidance helps preserve structural boundaries and reduce inaccurate context propagation across object boundaries.Source: B. Dual-Domain Feature Interaction… · pp. 5–6
  • GSR refines amplitude and phase spectra to adjust global exposure while preserving the low-frequency representation’s spatial structure.The amplitude spectrum relates to global intensity and exposure, while the phase spectrum retains spatial structure.Source: B. Dual-Domain Feature Interaction… · pp. 5–6
  • MaIR uses nested S-shaped scanning to model spatially varying illumination and suppress sensor noise across spatial regions.The scanning strategy propagates context between illumination regions while preserving local continuity.Source: Figure 4 · p. 5, B. Dual-Domain Feature Interaction… · p. 6

2) Cross-Frequency Enhancement (CFE) Block:

CFE addresses the challenge of refining noise-sensitive high-frequency details that depend on low-frequency illumination cues. It jointly refines directional high-frequency features and uses enhanced low-frequency context to guide their reconstruction.

2) Cross-Frequency Enhancement (CFE) Block:

  • Existing cross-frequency approaches mainly use low-to-high guidance and do not explicitly model relationships among directional high-frequency sub-bands.The passage notes that high-frequency details are noise-sensitive and depend strongly on low-frequency illumination cues.Source: 2) Cross-Frequency Enhancement (CFE)… · p. 6
  • CFE jointly refines the three directional high-frequency features and uses recovered low-frequency context to guide detail restoration.This is designed to promote cross-directional information interaction and structural consistency.Source: 2) Cross-Frequency Enhancement (CFE)… · p. 6
  • CFE aggregates directional high-frequency features with SKFF, then refines the fused representation using enhanced low-frequency context.The guided representation is further processed and split into three directional high-frequency components.Source: 2) Cross-Frequency Enhancement (CFE)… · p. 6

C. Loss Function

Training combines VAE pretraining losses with full-stage reconstruction and multi-scale consistency losses. The Charbonnier term supports illumination and color consistency at coarse scales and structure and detail refinement at finer scales.

C. Loss Function

  • VAE pretraining uses MSE and KL-divergence losses to learn multi-scale image representations while preserving a compact latent representation.The KL-divergence weight λKL is 10^-6.Source: C. Loss Function · p. 6
  • Full-stage training aligns the enhanced output with its ground truth using an ℓ1 loss.The reconstruction target is Igt for output Iout.Source: C. Loss Function · pp. 6–7
  • Multi-scale Charbonnier loss enforces coarse-scale illumination and color consistency while refining local structures and details at finer scales.It uses four scales, and λchar is set to 0.2 by default.Source: C. Loss Function · p. 7
  • The full-training objective combines the ℓ1 and Charbonnier losses.The section presents this combined objective as the overall full-training loss.Source: C. Loss Function · p. 7

A. Experimental Setup

BiFreqNet is evaluated on two UHD and two conventional-resolution LLIE benchmarks, with comparisons against 18 methods using PSNR and SSIM. The setup uses a frozen pretrained VAE and four-scale feature modeling.

A. Experimental Setup

  • The model uses four feature scales, stacked DFI modules, and a frozen VAE backbone that provides a stable generative prior.The channel dimensions are {16, 32, 64, 64}; DFI counts are {1, 1, 1, 2}.Source: A. Experimental Setup · p. 7
  • BiFreqNet is evaluated on UHD-LL, UHD-LOL4K, LOLv1, and LOLv2-Real, with the latter described as containing more challenging real low-light images.The datasets include paired training and test sets, with sizes specified for each benchmark.Source: A. Experimental Setup · p. 7
  • Comparisons cover 18 state-of-the-art methods, trained and tested on the same dataset partitions except GPP-LLIE, whose results come from its original paper.The comparison set includes general and UHD enhancement and restoration methods.Source: A. Experimental Setup · p. 7
  • Evaluation uses PSNR for pixel-level reconstruction fidelity and SSIM for structural similarity, with higher values indicating better quality.Source: A. Experimental Setup · p. 7
  • Visual comparisons report more faithful illumination and clearer details for BiFreqNet on UHD-LL and UHD-LOL4K.Source: Figure 5 · p. 8, Figure 6 · p. 8

B. Results on UHD Low-Light Image Enhancement

On UHD-LL and UHD-LOL4K, BiFreqNet reports strong restoration results, with its best metric depending on the dataset. Visual comparisons describe more uniform illumination and clearer structures and textures.

  • On UHD-LL, BiFreqNet reaches 27.39 dB PSNR and ties for best SSIM at 0.929; it also exceeds Reti-Diff by 0.47 dB PSNR and 0.024 SSIM.It improves SSIM over Wave-Mamba from 0.913 to 0.929, and exceeds UHDFour and UHDDIP in PSNR by 1.16 dB and 0.64 dB, respectively.Source: B. Results on UHD Low-Light Image… · p. 8
  • On UHD-LOL4K, BiFreqNet achieves the highest SSIM at 0.992, while its PSNR is 0.14 dB below Wave-Mamba and 0.04 dB below LLFormer.Against DiffLL, PSNR rises from 36.95 dB to 37.29 dB and SSIM from 0.989 to 0.992.Source: B. Results on UHD Low-Light Image… · pp. 8–9
  • Visual comparisons show more uniform illumination and clearer text, boundaries, and local textures than competing methods on the UHD datasets.The described examples include blurred or weakened structures, illumination inconsistency, and color deviations in other methods.Source: B. Results on UHD Low-Light Image… · p. 9, Figure 7 · p. 9

C. Results on General Low-Light Image Enhancement

On the conventional-resolution LOLv1 and LOLv2-Real benchmarks, BiFreqNet reports the best PSNR and SSIM among compared methods. Visual comparisons describe balanced illumination and preserved fine structures.

D. Ablation Study

Ablations on UHD-LL examine the DFI components and S2LFE design. The results support complementary low- and high-frequency processing and a balanced multi-scale block allocation.

1) Effect of DFI Module:

  • With full DFI, PSNR/SSIM reach 27.39 dB/0.929; S2LFE provides the major restoration capability, while CFE adds detail refinement with 0.69M additional parameters.Removing both blocks yields 22.81 dB PSNR and 0.784 SSIM; retaining only CFE or S2LFE increases PSNR by 2.23 dB or 4.22 dB, respectively.Source: 1) Effect of DFI Module · p. 10, Table III · p. 10, Figure 9 · p. 10

2) Internal Analysis of S2LFE Block:

3) Internal Analysis of CFE Block:

CFE refines high-frequency details using enhanced low-frequency context, and ablations support guidance in both directions. The paper also identifies color restoration in extremely low-light scenes as a limitation.

3) Internal Analysis of CFE Block:

4) Effect of the Bidirectional Guidance Mechanism :

V. LIMITATIONS

  • In extremely low-light scenes, insufficient RGB color information can leave white balance and color temperature inaccurate despite recovered illumination and scene content.The authors suggest physics-based camera spectral-sensitivity and illuminant-chromaticity priors as possible future remedies.Source: §V · pp. 11–12, Figure 14 · p. 12

VI. CONCLUSION

  • Across four general and UHD LLIE benchmarks, BiFreqNet performs favorably against existing state-of-the-art general and UHD LLIE methods.Source: §VI · p. 12

Related papers

Loading 2610.33402v1…