Retinex-inspired Unrolling with Cooperative Prior Architecture Search for Low-light Image Enhancement

Motivation

• Low-light images exhibit poor visibility, low contrast, and unexpected noise, motivating enhancement before downstream vision tasks.The affected applications include object detection, segmentation, and recognition.Source: §1 · p. 1

Limitations of Existing Methods

• Classical Retinex-based methods use handcrafted priors that may be insufficiently adaptive and can produce noise or over- and under-enhancement.Source: §1 · p. 1
• CNN-based methods often require elaborate architectures and carefully selected paired or unpaired training data, yet can perform unsatisfactorily on noisy real-world images.The cited reason is that their underlying models lack principled physical constraints for capturing intrinsic low-light structure.Source: §1 · pp. 1–2

RUAS

• RUAS uses Retinex-informed optimization models and unrolls their processes to establish a holistic enhancement-network propagation structure.The framework is designed to exploit latent low-light structures in real-world noisy scenarios.Source: §1 · p. 2
• RUAS automatically discovers embedded prior architectures through a cooperative bilevel strategy for illumination estimation and noise removal without paired or unpaired supervision during search.Source: §1 · p. 2

Results

• RUAS produces memory- and computation-efficient networks that perform favorably against state-of-the-art approaches and can search priors for different low-light scenarios.The paper points to Figure 1 for the comparative evaluation.Source: §1 · p. 2

CNN-based LIE

• CNN-based low-light enhancement methods include autoencoders, attention-enhanced U-Nets, frequency-based networks, Retinex decompositions, bilateral learning, self-supervision, zero-reference curves, and recursive bands.Source: §2.1 · p. 2
• These methods use varied supervision regimes, including paired, unpaired, self-supervised, zero-reference, and semi-supervised training.Source: §2.1 · p. 2

Neural Architecture Search

• Neural architecture search has explored evolutionary algorithms and reinforcement learning to generate or optimize network architectures and parameters.Source: §2.2 · p. 2
• Evolutionary and reinforcement-learning NAS approaches require a large amount of computation.Source: §2.2 · p. 2

Framework overview

• RUAS unrolls Retinex-inspired optimization models into a holistic enhancement network, followed by a distillation-cell search space and cooperative bilevel architecture search.The framework targets illumination estimation and noise removal while integrating optimization structure with learned priors.Source: §3 · p. 2, §3.1 · pp. 2–3, §3.2 · p. 3

Retinex-inspired optimization unrolling

• The Retinex rule models the observed image y as the element-wise product of the desired recovery x and illumination map t.IEM estimates t, while NRM suppresses noise in challenging low-light scenarios.Source: §3.1 · pp. 2–3

Illumination Estimation Module

• IEM warm-starts illumination from u_k = y ⊘ t_k, uses local-channel maxima and residual correction, then refines t across K unrolled stages.The residual term adaptively suppresses overexposed pixels, and the CNN parameterizes the regularization gradient in each stage.Source: §3.1.1 · p. 3

Noise Removal Module

• NRM solves a regularized model anchored to the IEM output u_K and unrolls updates whose regularization gradient is parameterized by CNRM.The module produces the recovered image x_N after N stages.Source: §3.1.2 · p. 3

Compact search space

• The compact search space represents CIEM and CNRM as five-node distillation cells with convolution, residual, dilation, and skip operations encoded by architecture vectors α.Architecture parameters and layer weights are jointly learned over mixed operations.Source: §3.2.1 · p. 4

Reference-free bilevel learning

• Cooperative bilevel learning alternates updates for illumination and noise-removal architectures using reference-free training and validation losses.The validation objective combines IEM and NRM losses, with the NRM task depending on the illumination architecture.Source: §3.2.3 · pp. 4–5, §3.2.2 · p. 4

Experiments

• Table 1 reports PSNR and SSIM comparisons on MIT-Adobe 5K and LOL, while Figure 3 visualizes DarkFace comparisons with red boxes marking obvious differences.The supplied figure and table passages identify the evaluation layouts but do not provide the individual numerical table entries.Source: Table 1 · p. 5, Figure 3 · p. 5

• RUAS obtained the best scores on the MIT-Adobe 5K benchmark among the evaluated methods.Source: §4.2 · p. 6
• On LOL, RUAS produced vivid colors and removed undesired noise, while competing methods introduced artifacts or failed to restore realistic colors.Source: §4.2 · p. 7
• RUAS achieved strong detail restoration and noise removal on extremely challenging real-world examples, where competing methods produced visible artifacts and color distortions.Source: §4.2 · p. 7
• RUAS required very small model size, FLOPs, and inference time compared with manually designed CNN-based methods.The comparison reports model size, FLOPs, and GPU-seconds on 100 LOL images sized 600×400.Source: §4.2 · p. 7, Table 2 · p. 7
• Residual rectification in the warm-start strategy suppressed over-exposure and produced more comfortable exposure than alternative strategies.Source: §4.3 · pp. 7–8
• NRM improved noisy-scene enhancement by removing visible noise that damaged results without the module.Source: §4.3 · p. 8
• The searched architecture achieved the highest PSNR with less inference time than naively designed alternatives on MIT-Adobe 5K.Source: §4.3 · p. 8, Figure 8 · p. 8
• Cooperative search connected illumination estimation and denoising, with the searched NRM containing more residual convolutions and skip connections.Source: §4.3 · p. 8

• RUAS integrates principled optimization unrolling with cooperative prior architecture search to design effective, lightweight low-light enhancement networks.Source: §5 · p. 8
• Retinex-based optimization models formulate latent illumination and desired-image structures, while unrolling produces the enhancement network’s holistic structure.Source: §5 · p. 8
• A cooperative reference-free strategy discovers specific architectures from a compact search space for low-light enhancement.Source: §5 · p. 8
• Experiments on challenging benchmark datasets produced new state-of-the-art results.Source: §5 · p. 8