Frequency-Based LLIE Methods

Updated 24 March 2026

Frequency-based LLIE methods are techniques that leverage spectral decompositions (Fourier, Laplace, wavelet) to separate global illumination from fine details in low-light images.
They utilize explicit transforms to isolate low-frequency brightness from high-frequency textures and noise, enabling targeted restoration and enhanced perceptual fidelity.
These methods integrate frequency-spatial fusion mechanisms and multi-scale loss formulations to achieve efficient, high-definition image restoration across diverse domains.

Frequency-based low-light image enhancement (LLIE) methods leverage spectral representations—Fourier, wavelet, and Laplace decompositions—to disentangle illumination and detail, yielding more effective global brightness and fine-structure restoration in severely underexposed images. This paradigm exploits the decoupling of global illumination (typically encoded in low-frequency components) from high-frequency detail (textures, edges, or noise) to achieve improvements in both perceptual fidelity and restoration accuracy, with efficient computational regimes and strong adaptability to ultra-high-resolution and domain-specific tasks.

1. Spectral Decomposition Frameworks

Frequency-based LLIE methods rely on explicit transforms to project input images or features into spectral domains, exposing frequency-localized content for targeted processing.

Fourier Transform: Used in "FourLLIE" (Wang et al., 2023) and "FSIDNet" (Tao et al., 25 Oct 2025), the 2D DFT decomposes an image $x(h,w)$ into amplitude $A(u,v)$ (energy) and phase $\Phi(u,v)$ (structure). Amplitude correlates with global brightness, while phase encodes edges.
Laplace Pyramid: "AFD-LLIE" (Zhou et al., 2024) employs a $K$ -level Laplacian decomposition, recursively generating multi-scale high-frequency detail and low-frequency (illumination) maps. Reconstruction (Eq. (2)) inverts the pyramid, ensuring lossless multi-frequency aggregation.
Wavelet Transform: Both "Wave-Mamba" (Zou et al., 2024) and "LLCaps" (Bai et al., 2023) utilize DWT to split image content into approximation (LL band, dominant content) and detail bands (LH, HL, HH, noise/textures), supporting exact downsampling and sub-band manipulation across scales.

These frameworks enable precise targeting of illumination (low-frequency) and detail/noise (high-frequency) components, often motivating staged or multi-branch network architectures for their independent restoration.

2. Decoupling Illumination and Detail Restoration

A defining property of frequency-based LLIE is the explicit separation of global illumination recovery from fine detail enhancement.

Amplitude-Phase Separation (Fourier Domain): Both "FourLLIE" (Wang et al., 2023) and "FSIDNet" (Tao et al., 25 Oct 2025) first correct amplitude for global lightness, then refine details through phase restoration or spatial-frequency fusion. In "FSIDNet", amplitude restoration employs $\hat Y_1 = \mathcal{F}^{-1}\{\hat A,\,e^{j\Phi(X_{in})}\}$ , while phase restoration applies $\hat Y_2 = \mathcal{F}^{-1}\{\hat A,\,\hat\Phi\}$ .
Laplace Disentanglement (AFD-LLIE): The "coarse" stage recovers low-frequency maps (illumination), followed by a "fine" Laplace Decoupled Restoration Model that reconstructs all Laplace scales with explicit low-frequency consistency loss, as formalized in Eq. (3) (Zhou et al., 2024).
Wavelet and Curved Attention: "Wave-Mamba" processes LL bands through state space models for content, while high-frequency enhance blocks refine details by injecting processed LL information (Zou et al., 2024). "LLCaps" leverages curved wavelet attention to amplify high-frequency edges and localized illumination (Bai et al., 2023).
Frequency Illumination Adaptor (RetinexDual): Illumination is globally corrected with a Fourier-convolutional branch, while reflectance restoration remains spatial, avoiding adverse effects of frequency-domain processing on fine local artifacts (Kishawy et al., 6 Aug 2025).

This separation avoids the mutual interference between global lighting and intricate detail that is characteristic of spatial, “one-shot” LLIE networks.

3. Spectral-Interaction and Fusion Mechanisms

Advanced frequency-based LLIE frameworks integrate frequency and spatial cues using tailored fusion mechanisms to ensure complementary information exchange.

Frequency-Spatial Interaction Blocks: FSIA/FSIP blocks (FSIDNet (Tao et al., 25 Oct 2025)) process features in parallel spatial and frequency branches, with cross-branch fusion via convolutional mixing of Fourier-processed and spatial outputs.
Signal-to-Noise Ratio Guided Fusion: "FourLLIE" (Wang et al., 2023) computes an SNR map to modulate per-pixel fusion of local CNN and global Fourier features, emphasizing reliable spatial detail and suppressing noise in ambiguous regions.
Information Exchange Modules: FSIDNet's IEM propagates cross-scale and cross-stage features, with spatially adaptive filtering to preserve low-level cues and ensure holistic restoration across network depth (Tao et al., 25 Oct 2025).
LL-Guided HF Correction: "Wave-Mamba" injects enhanced low-frequency features into high-frequency refinement (via FMT), correcting detail in a content-aware manner (Zou et al., 2024).
Curved Wavelet Attention: LLCaps applies local nonlinear curve-based attention within each wavelet sub-band, boosting adaptability to varying illumination distributions (Bai et al., 2023).

These mechanisms underpin the superior capability of frequency-based LLIE methods to simultaneously handle global and local degradations.

4. Loss Formulations and Multi-Scale Supervision

Loss formulations in frequency-based LLIE often directly supervise both frequency and spatial outputs, yielding stronger and more stable gradients.

Multi-Scale or Multi-Component Loss: "AFD-LLIE" applies multi-scale $\ell_1$ loss across all Laplace bands plus a low-frequency consistency term (Eq. (3)), enforcing band-specific optimization (Zhou et al., 2024).
Frequency-Domain Objectives: "RetinexDual" deploys combined Charbonnier, frequency-domain $\ell_1$ , SSIM, and VGG perceptual losses at all three decoding scales, coupling both spatial and spectral fidelity (Kishawy et al., 6 Aug 2025).
Wavelet and Amplitude Supervision: LLCaps includes a wavelet-domain reconstruction loss and Charbonnier loss, while "FourLLIE" optimizes amplitude spectra directly in the frequency stage, followed by perceptual and spatial reconstruction in later stages (Wang et al., 2023, Bai et al., 2023).
End-to-End Pixel + Spectral Losses: "Wave-Mamba" requires no sub-band loss and supervises the final output via single-scale $\ell_1$ (Zou et al., 2024), whereas "FSIDNet" explicitly supervises amplitude and phase restoration sequentially with distinct weights for amplitude- and phase-domain error (Tao et al., 25 Oct 2025).

These multifaceted objectives increase convergence stability and accentuate desired behaviors in both global correction and texture recovery.

5. Quantitative Outcomes and Efficiency

Frequency-based LLIE methods substantiate their effectiveness through marked improvements in both restoration accuracy and computational/resource efficiency.

Method	Backbone/Domain	Peak PSNR Gains/Performance	Param. Overhead	Efficiency Notable
AFD-LLIE (Zhou et al., 2024)	MIR-Net, Restormer, etc.	+7.68 dB (SDSD-out, Restormer); +4.62 dB (LOL-v2)	+88 K (0.2–5.5%)	+2.53 GFLOPS, +8ms (256²)
FourLLIE (Wang et al., 2023)	Paired LLIE datasets	PSNR 22.34–24.65 dB (SOTA)	0.12 M (lightweight)	~40 fps @ 384²
FSIDNet (Tao et al., 25 Oct 2025)	LOL-Real, LSRW-Huawei	PSNR 22.65 dB, SSIM 0.852 (LOL)	Not listed	Efficient U-Net, skip links
Wave-Mamba (Zou et al., 2024)	UHD LLIE	PSNR 37.43 dB (UHD-LOL4K), 27.35 dB (UHD-LL)	1.26 M (linear time)	Real-time 4K, preserves info
LLCaps (Bai et al., 2023)	Capsule Endoscopy	+3.57 dB (Kvasir, vs. MIRNetv2)	Not specified	Wavelet + diffusion
RetinexDual (Kishawy et al., 6 Aug 2025)	UHD-LLIE	PSNR +2.12 dB attributable to FIA	Parallel FIA/SAMBA	Lightweight on illumination

A notable result is that frequency disentanglement methods often deliver marked boosts in PSNR (up to +7.68 dB), with marginal increases in parameter count or inference time. Single-block or spectral-branch overheads remain small relative to backbone network capacity, and explicit frequency manipulation, as in DWT or FFT branches, confers substantial performance gains even in ultra-high-definition or highly domain-constrained scenarios.

6. Comparative Advantages and Limitations

Frequency-based LLIE methods outperform traditional spatial or naïvely frequency-augmented networks due to:

Targeted, Decoupled Optimization: Separate modules for illumination and detail avert interference and extract better global context (Zhou et al., 2024, Tao et al., 25 Oct 2025, Kishawy et al., 6 Aug 2025).
Global Context Accessibility: FFT/DWT manipulations inherently leverage the entire image content, enabling efficient correction of global color and exposure distributions (Kishawy et al., 6 Aug 2025).
Efficient Parameterization: 1×1 convolutions or lightweight SSMs in the spectral domain offer content-adaptive processing with reduced computational budgets compared to competing approaches (e.g., Transformers) (Wang et al., 2023, Zou et al., 2024).
Exact Downsampling and Information Retention: DWT/IDWT used in "Wave-Mamba" and "LLCaps" avoids information loss linked to pooling or strided CNNs, vital for UHD image restoration (Zou et al., 2024, Bai et al., 2023).

However, these methods are less effective when attempting to restore highly localized degradations solely via frequency operations, as shown by the drop in PSNR and SSIM when FIA is misapplied to fine-detail restoration in RetinexDual (Kishawy et al., 6 Aug 2025). Purely frequency-based models may also struggle with artifacts confined to small spatial support, necessitating spatial or hybrid branches for full restoration.

7. Future Directions and Domain Adaptations

Recent work extends frequency-based LLIE to broader restoration domains:

Adaptation to UHD and Real-Time Applications: Linear-complexity SSMs and exact wavelet downsampling are enabling frequency-based LLIE in 4K video and medical imaging (Zou et al., 2024, Bai et al., 2023).
Domain-Specific Attention Mechanisms: Curved wavelet attention addresses domain-tailored illumination priors in endoscopy (Bai et al., 2023).
Modular Integration: Wrapper-style frequency modules (e.g., ACCA + Laplace decoupling (Zhou et al., 2024), RetinexDual’s FIA (Kishawy et al., 6 Aug 2025)) facilitate straightforward enhancement of diverse backbones (CNNs, Transformers, diffusion models) without architectural overhaul.
Hybrid Decomposition: Future research will likely pursue tighter synergy between spatial, Laplace, Fourier, and wavelet decompositions, exploiting strengths of each for multiscale, multimodal restoration.

These advances suggest a continuing shift towards spectral-leveraged, computationally efficient, and domain-adaptive image enhancement models. The consistent, large performance gains and rapid inference achieved by frequency-based LLIE establish these methods as essential for high-fidelity restoration in demanding lighting and resolution conditions.