Illumination-Guided Event Enhancement Module

Updated 12 January 2026

The paper introduces a module that fuses intensity images and event data using illumination-guided adaptive mechanisms to enhance low-light image restoration.
It leverages dual-domain fusion by combining spatial and Fourier-domain filtering with deformable convolution and cross-modal attention to align features.
The module achieves state-of-the-art performance with improved PSNR and SSIM, effectively recovering details and reducing noise under severe low-light conditions.

An Illumination-Guided Event Enhancement Module (IG-EEM) is an architectural component designed to leverage both frame-based intensity images and event camera data for robust restoration and enhancement under severe low-light conditions. IG-EEM variants orchestrate cross-modal fusion and dual-domain adaptation—often with explicit illumination guidance—to suppress noise, recover structure, and maximize visual fidelity. This approach is increasingly fundamental in modern low-light image enhancement networks, particularly as event cameras enable microsecond-level sensing and high dynamic range. IG-EEMs typically coordinate the interaction of spatial, spectral, and temporal modalities using attention, deformable alignment, Fourier-domain filtering, and illumination-adaptive mechanisms.

1. The Role and Rationale of Illumination-Guided Enhancement

Illumination-Guided Event Enhancement addresses challenges posed by extreme lighting, noise amplification, and spatial-temporal misalignment. Traditional intensity-frame methods, prone to underexposure and blur, fail to retain critical structure or color fidelity. Event cameras, which asynchronously record pixel-wise brightness changes, provide high dynamic range and temporal resolution but are susceptible to flicker artifacts and low-frequency noise under abrupt illumination changes.

IG-EEM modules inject illumination cues—often estimated from the low-light frame and event responses—into subsequent event-image fusion stages. This guidance is realized via explicit per-channel illumination scaling, spatially local and global attention blocks, SNR-adaptive feature selection, or spectral filtering. The net effect is to dynamically modulate both the aggressiveness of denoising filters and the alignment of multimodal features, thus ensuring that enhancement respects spatially varying illumination and maximizes recovery of details and contrast (Cai et al., 1 Jul 2025, Liu et al., 6 Jun 2025, Liu et al., 2024, Guo et al., 4 Mar 2025, Sun et al., 13 Apr 2025, Kai et al., 5 Jan 2026, Chen et al., 2024).

2. Architectural Mechanisms and Data Flow

IG-EEM implementations vary but commonly combine these core operations:

Illumination Estimation: A shallow CNN or U-Net computes a rough or precise illumination map Ĺ or L from the low-light frame and/or event voxel grid. This map may be driven by classical Retinex theory—factorizing the image into reflectance and illumination components for targeted enhancement (Guo et al., 4 Mar 2025, Sun et al., 13 Apr 2025, Wu et al., 30 Nov 2025, Kai et al., 5 Jan 2026, Chen et al., 2024).
Dual-Domain Fusion: Features are split or processed in both spatial and frequency domains. Frequency branches typically use adaptive Gaussian filtering or high-pass Fourier-domain filters with channel-wise scales σ_c modulated by illumination cues; spatial branches deploy deformable convolutions where offsets are illumination-dependent (Cai et al., 1 Jul 2025, Liu et al., 6 Jun 2025, Kai et al., 5 Jan 2026).
Cross-Modal Attention: Joint fusion employs cross-attention mechanisms that exchange keys/values (often learned) between image and event modality tokens, further adapted by learned gates or explicit SNR-maps (Cai et al., 1 Jul 2025, Wu et al., 30 Nov 2025, She et al., 1 Aug 2025, Chen et al., 2024).
Hierarchical Attention Fusion: Outputs of spatial and frequency branches are fused by spatial and channel attention modules, sometimes realized as fully hierarchical multi-scale cascades (Cai et al., 1 Jul 2025, Kai et al., 5 Jan 2026, Chen et al., 2024).
SNR-Adaptive Region Selection: SNR maps estimate local reliability, guiding fusion such that high-SNR (well-lit) image regions are preferred, while low-SNR (dark, noisy) regions receive structural cues from events (Wu et al., 30 Nov 2025, Chen et al., 2024).
Temporal Enhancement: ConvGRU or convLSTM units are inserted to ensure temporal coherence and memory persistence, further regularized by temporal loss terms (Liu et al., 2024, Chen et al., 2024).
Final Decoding: The enhanced features are projected via convolution layers to synthesize the output image with improved contrast, noise suppression, and restored spatial detail (Cai et al., 1 Jul 2025, Guo et al., 4 Mar 2025, Chen et al., 2024).

3. Core Mathematical Formulations

Typical IG-EEM modules instantiate several advanced computational blocks:

Adaptive Frequency Filtering:

$G_{σ_c}(x, y) = \frac{1}{2πσ_c^2} \exp\left(-\frac{x^2 + y^2}{2σ_c^2}\right)$

$Ŷ_{freq, c} = \mathcal{F}^{-1}\left(\mathcal{F}(X_{freq, c}) \cdot \mathcal{F}(G_{σ_c})\right)$

Used in both (Cai et al., 1 Jul 2025) (adaptive Gaussian) and (Liu et al., 6 Jun 2025) (high-pass filter via Fourier domain).

Deformable Convolution Alignment:

$Ŷ_{spat}(p_0) = \sum_{k=1}^K w_k \cdot X_{spat}(p_0 + p_k + Δp_k)$

with $Δp_k$ learned from illumination-aware features (Cai et al., 1 Jul 2025, She et al., 1 Aug 2025).

Cross-modal RWKV Attention:

$wkv_t = \frac{ \sum_{i≠t} e^{-\frac{|t-i|-1}{T}(w + k_i)} v_i + e^{u + k_t} v_t }{ \sum_{i≠t} e^{-\frac{|t-i|-1}{T}(w + k_i)} + e^{u + k_t} }$

employed for continuous spatial-temporal fusion (Cai et al., 1 Jul 2025).

SNR-Guided Region Selection:

$M_{snr}(x) = \frac{\mu_g(x)}{ |I_g(x) - \mu_g(x)| + ε }$

$F_{sel-img} = \hat{M}_{snr} \odot \hat{F}_{img}$

$F_{sel-ev} = (1 - \hat{M}_{snr}) \odot \hat{F}_{ev}$

for regional fusion, as in (Wu et al., 30 Nov 2025, Chen et al., 2024).

Loss Functions:

Multi-term objectives integrating reconstruction loss (Charbonnier or L₁), perceptual loss (AlexNet or VGG), SSIM, multi-scale SSIM, frequency loss, and temporal consistency loss:

$\mathcal{L}_{total} = λ_r\mathcal{L}_r + λ_p\mathcal{L}_p + λ_s\mathcal{L}_s + λ_m\mathcal{L}_m$

as in (Cai et al., 1 Jul 2025, Liu et al., 6 Jun 2025, Liu et al., 2024, She et al., 1 Aug 2025, Chen et al., 2024).

4. Comparative Implementations and Performance

Numerous architectures employ IG-EEM modules but differ in their fusion strategies and adaptation mechanisms:

Method	Fusion Modality	Illumination Guidance	Key Mechanisms
EvRWKV (Cai et al., 1 Jul 2025)	Image/Event frequency/spatial	Learned σ_c, rough Ĺ, event voxel	Cross-RWKV, FFT Gaussian, Deform
BiLIE (Liu et al., 6 Jun 2025)	Event spectral	Gaussian HPF (σ=12)	FFT HPF, no learnable params
ERetinex (Guo et al., 4 Mar 2025)	Image/Event branch	U-Net fusion	Light-up L^{-1}
RetinEV (Sun et al., 13 Apr 2025)	Event-derived illumination	Temporal-mapping events	T2I network, cross-modal attn
RetinexEVSR (Kai et al., 5 Jan 2026)	Event/illumination multi-scale	SCI-derived I_t	Multi-scale feature fusion
EvLight++ (Chen et al., 2024)	Image/Event regional	SNR-adaptive selection	IRFS/ERFS, Holistic fusion

Experimental results strongly validate IG-EEMs. EvRWKV+EISFE achieves state-of-the-art PSNR/SSIM across SDE, SDSD, and RELED datasets, improving over previous solutions by at least 0.8 dB (Cai et al., 1 Jul 2025). EvLight++'s SNR-guided fusion shows substantial gains in both video enhancement and downstream segmentation tasks (Chen et al., 2024). Retinex-inspired methods (ERetinex, RetinEV, RetinexEVSR) demonstrate superior detail recovery at far lower computational cost and sizable PSNR improvements (Guo et al., 4 Mar 2025, Sun et al., 13 Apr 2025, Kai et al., 5 Jan 2026).

5. Illumination-Guided Adaptation: Regional, Spectral, and Temporal Perspectives

A defining technical advance is spatially adaptive enhancement. SNR mapping enables region-wise selection: in well-lit regions, image features dominate; in noise-challenged or underexposed regions, event-derived edges or structure are prioritized (Wu et al., 30 Nov 2025, Chen et al., 2024, Sun et al., 13 Apr 2025). Frequency-domain filtering, modulated by per-channel illumination scale, improves denoising and edge recovery without amplifying low-frequency global flicker (Cai et al., 1 Jul 2025, Liu et al., 6 Jun 2025).

Temporal guidance is achieved with convLSTM/ConvGRU blocks, regularizing feature propagation and enforcing consistency over time. Nighttime-specific variants insert non-uniform illumination adaptation at each encoding stage, suppressing artifacts and stabilizing event streams subjected to highly variable lighting (Liu et al., 2024).

6. Cross-Domain Applicability and Emerging Trends

Event-driven illumination guidance generalizes across imaging domains. In traffic object detection, AFCM modules use global lightness distribution from images to adaptively gate event features via attention, yielding robust detection under variable illumination (Liu et al., 2023). For 3D geometry estimation, SNR-aware fusion mechanisms enable depth/pose prediction without retraining under night conditions, and photometric event losses reinforce consistency in spatiotemporal optimization (Wu et al., 30 Nov 2025).

Recent designs rely increasingly on multi-scale, hierarchical fusion across both spatial and frequency domains, integrating learning-based adaptation with prior-driven filtering. Auxiliary modules borrow Retinex-derived priors or synthetic event signals for broader dynamic range and lighting condition flexibility (Guo et al., 4 Mar 2025, Sun et al., 13 Apr 2025, Kai et al., 5 Jan 2026, LU et al., 28 Feb 2025).

7. Experimental Validation and Impact

Quantitative and qualitative ablation studies consistently demonstrate the performance benefits of IG-EEM integration. Notable improvements include:

EvRWKV+EISFE: PSNR 23.09 dB (SDE-in), 28.96 dB (SDSD-in), 32.18 dB (RELED); consistently higher SSIM relative to SOTA (Cai et al., 1 Jul 2025).
BiLIE+EFE: PSNR improvement from 19.69→20.82 dB on the RELIE dataset; visual sharpening of edges and suppression of illumination-induced flicker (Liu et al., 6 Jun 2025).
RetinEV: up to 6.62 dB PSNR improvement on real low-light datasets with efficient frame rate (Sun et al., 13 Apr 2025).
EvLight++: Outperforms single-image/video methods by 1.37–3.71 dB and boosts semantic segmentation mIoU by 15.97% over enhanced outputs (Chen et al., 2024).

IG-EEM modules are now critical in event-guided low-light imaging pipelines, enabling consistent structural detail recovery, superior noise suppression, and robust adaptation across real-world, synthetic, and dynamically illuminated scenarios.

References:

EvRWKV (Cai et al., 1 Jul 2025), BiLIE (Liu et al., 6 Jun 2025), SEE-Net (LU et al., 28 Feb 2025), ERetinex (Guo et al., 4 Mar 2025), RetinEV (Sun et al., 13 Apr 2025), RetinexEVSR (Kai et al., 5 Jan 2026), EAG3R (Wu et al., 30 Nov 2025), SFNet (Liu et al., 2023), EvLight++ (Chen et al., 2024).