Low-Light Image Enhancement (LLIE)

• Low-light image enhancement (LLIE) is a computational process that improves images captured in insufficient light by restoring brightness, texture, and color fidelity.
• Recent methods combine classical models like Retinex with frequency-domain processing and deep architectures (CNNs, transformers) to mitigate noise and enhance details.
• Efficient, task-aware techniques enable real-time, edge-deployable enhancements that balance visual quality for humans with performance in automated vision tasks.

Low-light image enhancement (LLIE) is the computational process of improving the perceptual quality and information content in images degraded by insufficient illumination. LLIE addresses fundamental computer vision and imaging challenges, including visibility restoration, texture/detail retrieval, denoising, and photometric/color correction. State-of-the-art LLIE methods now exploit both signal processes (illumination modeling, frequency analysis, denoising) and advanced machine learning architectures (deep CNNs, transformers, diffusion models, reinforcement learning) to robustly handle a wide range of real-world degradations while balancing computational efficiency.

1. Physical Models and Fundamental Signal Properties

Classical LLIE methods are typically grounded in the Retinex theory, which decomposes an observed low-light image $L(x)$ into reflectance $I(x)$ (scene content) and illumination $T(x)$ (lighting conditions): $L(x) = I(x) \circ T(x)$. The direct per-pixel decomposition, as instantiated in the LIME method, estimates $T(x)$ by local channel maxima and refines it with structure-aware regularization, enabling brightening without over-amplifying noise or inducing artifacts (Guo, 2016). However, such approaches are limited by the ill-posedness of factorization, which can entangle noise with reflectance or exaggerate color distortions.

Recent work identifies a crucial insight: in the frequency domain, the amplitude of the Fourier spectrum of the image encodes much of the perceived brightness, while the phase encodes structural details. Swapping amplitude spectra between low- and normal-light photos can transfer global exposure while holding structure fixed (Wang et al., 2023). This motivates frequency-space processing as a new model axis.

2. Architectural Advances: Frequency, Spatial, and Hybrid Domain Processing

Frequency-based LLIE methods (e.g., FourLLIE, FSIDNet) employ fast Fourier transform (FFT/IFFT) in tandem with learnable convolutional operations. In FourLLIE, stage one explicitly enhances the amplitude spectrum using a neural network-estimated amplitude transform map; the result is then fused with localized, spatially refined enhancements modulated by a signal-to-noise ratio (SNR) map, which distinguishes where global versus local processing is optimal (Wang et al., 2023). FSIDNet extends this paradigm, employing a two-stage U-Net: the first focuses on amplitude (lightness), the second on phase (structure), with information flow facilitated by bespoke frequency-spatial interaction blocks and a cross-stage Information Exchange Module (Tao et al., 25 Oct 2025).

Hybrid designs such as CIDNet leverage trainable color spaces—such as the HVI space that separately encodes intensity and orthogonal chromatic axes—to improve stability and mitigate color/brightness artifacts. A dual-branch architecture with cross-attention permits targeted denoising/inference in color and intensity channels (Yan et al., 2024).

Retinex-inspired deep models (e.g., LUMINA-Net, RLED-Net, LDE-Net) exploit data-driven variants of illumination-reflectance decomposition but mitigate classical pitfalls by operating in latent feature spaces, integrating channel/spatial attention, low-rank subspace noise suppression, and cross-branch/fusion mechanisms (Siddiqua et al., 21 Feb 2025, Ren et al., 2022, Zheng et al., 2024). Ablations consistently demonstrate that decoupling global and local correction, and separating signal and noise paths, strongly improves perceptual quality and robustness.

3. Learning Paradigms: Supervised, Unsupervised, and Reinforcement Techniques

While many LLIE systems require paired low-light/normal-light training sets, several lines of work circumvent this with unsupervised or weakly supervised strategies. Notable examples include:

• Self-calibrated denoisers combined with learnable illumination interpolation and self-regularization losses based on natural image statistics (Liu et al., 2023). Explicit noise estimation enables dynamic denoising, and a statistical regularizer (anchored on ImageNet means and variances) aligns the enhanced output's color and contrast to human expectations.
• Reinforcement learning approaches (e.g., ReLLIE) model LLIE as a Markov decision process where a fully-convolutional agent learns to iteratively apply per-pixel, adaptive enhancement curves. Reward functions integrate spatial consistency, exposure, smoothness, and channel-ratio loss, allowing direct control over enhancement strength and subjective preferences (Zhang et al., 2021).
• Feather-light models (FLIGHT-Net, LiteIE, SCLM) demonstrate that even extremely compact CNNs with minimal parameters—leveraging pixel-wise gain maps, iterative or local adaptation, and unsupervised losses—can rival much larger networks in both speed and perceptual quality, thereby enabling real-time, on-device enhancement (Ozcan et al., 2023, Zhang et al., 2023, Bai et al., 6 Jul 2025).

4. Incorporation of Priors and Advanced Generative Modeling

High-performing LLIE frameworks now integrate sophisticated priors, both from learned codebooks and generative models:

• Codebook- and quantization-based systems (LightQANet, CodeEnhance) pretrain vector-quantized autoencoders on well-lit reference datasets to provide high-fidelity discrete feature banks. Enhancement for low-light inputs is framed as learning to map into this codebook, with additional control via semantic embeddings or user-driven interactive feature transformation (Wu et al., 16 Oct 2025, Wu et al., 2024).
• Generative Perceptual Priors (GPP-LLIE) leverage off-the-shelf vision-LLMs to compute global and local perceptual quality priors (e.g., contrast, visibility, sharpness), which are then injected as conditioning signals in diffusion-based transformer networks. New normalization and attention mechanisms (GPP-LN, LPP-Attn) are designed to incorporate these priors at both global and pixel levels, delivering both state-of-the-art PSNR/SSIM and visually realistic results (Zhou et al., 2024).
• Event-based approaches (RetinEV) exploit temporal-mapping events from event cameras as direct illumination cues, using photodiode physics to recover pixel-wise exposure and integrating this signal into a reflectance-illumination decomposition pipeline. This yields significant PSNR gains over both image-only methods and traditional motion-event-based enhancement (Sun et al., 13 Apr 2025).

5. Comparison, Benchmarking, and Task-Driven Evaluation

LLIE benchmarking traditionally focuses on paired datasets (LOL, SICE, LSRW, LoLI-Street) for full-reference metrics (PSNR, SSIM, LPIPS) and unpaired sets (DICM, LIME, MEF, NPE, VV) evaluated by no-reference indicators (NIQE, BRISQUE) (Islam et al., 2024, Tao et al., 25 Oct 2025). TriFuse recently combined discrete wavelet transform, transformer-based diffusion modules, and edge-sharpening to achieve SOTA performance on LoLI-Street and generalization on diverse domains.

However, a major empirical finding is a weak or even negative correlation between human-centric LLIE metrics and machine-vision downstream accuracy. Detailed ablation on classification (VGG, ResNet, ViT) and detection (YOLOv7, EfficientDet) tasks reveals that enhancement approaches seeking maximal PSNR/SSIM or perceptual scores sometimes undermine semantic boundaries or color statistics relied upon by modern vision pipelines, thus degrading or failing to improve object detection/classification performance (Wu et al., 2024). This dichotomy suggests that "visually pleasing" LLIE for humans is often not aligned with task-optimized representations for machines.

6. Efficiency, Redundancy, and Model Design for Edge Deployment

Model redundancy—both parameter "harmfulness" (fixed weights that reduce performance on some samples) and "uselessness" (dynamic parameters that do not adapt)—has emerged as a key theme in LLIE engineering. Attention Dynamic Reallocation (ADR) and Parameter Orthogonal Generation (POG) have been proposed to remove such redundant weights and improve both accuracy and efficiency across transformer and U-Net backbones (Li et al., 2024). Lightest-weight architectures (LiteIE with 58 parameters, FLIGHT-Net at 25K parameters) now offer practical, real-time enhancement at 4K on mobile SoCs with negligible memory footprint (Bai et al., 6 Jul 2025, Ozcan et al., 2023), owing to model simplification, iterative restoration, and efficient loss formulations.

7. Future Research Directions and Open Challenges

Several trajectories are suggested for advancing LLIE:

• Task-aware LLIE: Aligning enhancement outputs with the semantic/objective needs of downstream tasks via perceptual and feature-space task losses, or through multi-modal guidance (e.g., CLIP embeddings) (Wu et al., 2024).
• Domain generalization: Extending models to real-world, non-paired data, handling extreme/heterogeneous illuminations (as in LoLI-Street and EvLowLight datasets), and transferring to RAW or event-based input regimes (Islam et al., 2024, Sun et al., 13 Apr 2025).
• Hybrid and multi-modal fusion: Further integrating spatial, frequency, event, codebook, and semantic streams to resolve the trade-off between global exposure, local detail, denoising, and color fidelity.
• Real-time and memory-constrained deployment: Continual advancement in parameter- and compute-efficient models, especially for embedded/edge applications, without sacrificing robustness or perceptual quality (Bai et al., 6 Jul 2025).
• Human-in-the-loop and customizable LLIE: Developing fully controllable pipelines (as in BCNet, CodeEnhance, ReLLIE), wherein users or high-level agents can direct the enhancement trajectory (saturation, style, scene priors).

LLIE remains an active field blending imaging physics, signal processing, deep learning, and deployment-driven design, with current research pushing the limits of generalization, efficiency, and functional fidelity.