[论文解读] Retinex-inspired Unrolling with Cooperative Prior Architecture Search for Low-light Image Enhancement

Risheng Liu, Long Ma+3 more|arXiv (Cornell University)|Dec 10, 2020

Image Enhancement Techniques参考文献 27被引用 59

一句话总结

RUAS 将 Retinex 启发的展开优化与协作式、无参考神经架构搜索结合，以构建轻量级、高性能的低光照图像增强网络。它在较低的计算成本下展示了最先进的结果。

ABSTRACT

Low-light image enhancement plays very important roles in low-level vision field. Recent works have built a large variety of deep learning models to address this task. However, these approaches mostly rely on significant architecture engineering and suffer from high computational burden. In this paper, we propose a new method, named Retinex-inspired Unrolling with Architecture Search (RUAS), to construct lightweight yet effective enhancement network for low-light images in real-world scenario. Specifically, building upon Retinex rule, RUAS first establishes models to characterize the intrinsic underexposed structure of low-light images and unroll their optimization processes to construct our holistic propagation structure. Then by designing a cooperative reference-free learning strategy to discover low-light prior architectures from a compact search space, RUAS is able to obtain a top-performing image enhancement network, which is with fast speed and requires few computational resources. Extensive experiments verify the superiority of our RUAS framework against recently proposed state-of-the-art methods.

研究动机与目标

• Motivate robust low-light image enhancement that preserves texture while reducing noise.
• Develop a principled network by unrolling Retinex-based optimization to model illumination and noise separately.
• Automatically discover compact prior architectures for illumination estimation and denoising without requiring paired data.
• Propose a cooperative bilevel learning framework to search and train lightweight, effective architectures.

提出的方法

• Formulate Retinex-based models with Illumination Estimation Module (IEM) and Noise Removal Module (NRM).
• Unroll optimization steps to build a propagation network where each iteration is parameterized by a CNN block.
• Define a compact search space via distillation cells with selectable operations (1x1/3x3 convs, residual convolutions, dilated convolutions, skip connections).
• Use differentiable NAS to obtain architecture parameters for IEM and NRM in a cooperative, bilevel learning setup.
• Employ reference-free losses for training: fidelity plus regularization terms (RTV for IEM, TV for NRM).
• Solve a cooperative min-min optimization over architecture parameters to jointly optimize IEM and NRM architectures.

实验结果

研究问题

• RQ1 Can Retinex-inspired unrolling yield a lightweight yet effective hierarchy for low-light enhancement?
• RQ2 Can a cooperative, reference-free NAS strategy identify compact architectures for illumination estimation and noise removal that outperform manually designed networks?
• RQ3 How does RUAS balance brightness, texture preservation, and noise suppression across real-world low-light scenarios?
• RQ4 What are the runtime and memory benefits of the RUAS-searched architectures compared to state-of-the-art CNNs?

主要发现

Table 1 – 数据集 / 指标 / 方法	PSNR	SSIM	MIT-Adobe 5K (PSNR)	MIT-Adobe 5K (SSIM)	LOL (PSNR)	LOL (SSIM)	Table 2 – 模型效率	大小 (M)	FLOPs (G)	时间 (S)	RUAS i	RUAS i+n
MIT-Adobe 5K – PSNR (dB)		LIME 17.788	SSIM 0.826	LIME 17.788	SSIM 0.826	RUAS i (PSNR) 20.830, SSIM 0.854	0.001
MIT-Adobe 5K – PSNR (dB)	SSIM	SDD 17.617	0.792	DRBN 15.954	0.704	RUAS i+n (PSNR) 20.830, SSIM 0.854	0.003
LOL – PSNR (dB)	LOL – SSIM	LIME 14.916	0.516	SDD 15.484	0.578	KinD 14.616	0.636	Ours 18.226	0.717	Table 2 values	0.003

• RUAS achieves state-of-the-art or competitive PSNR/SSIM on MIT-Adobe 5K and LOL datasets with significantly smaller model sizes and FLOPs.
• The cooperative search yields architectures for IEM and NRM that outperform separate or naive joint searches in PSNR, SSIM, and efficiency.
• Including the Noise Removal Module (NRM) improves performance in noisy, real-world low-light scenes.
• Searched RUAS variants (RUAS i and RUAS i+n) show superior efficiency (smaller size, lower FLOPs, faster inference) compared to several manually designed baselines.
• Ablation studies confirm the benefits of refined warm-start strategies and the necessity of NRM for noisy scenarios.