Retinex-inspired Unrolling with Cooperative Prior Architecture Search for Low-light Image Enhancement (2012.05609v1)

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.

• The paper introduces RUAS, a method that leverages Retinex theory and unrolling techniques to separate illumination and reflectance in low-light images.
• The paper employs a cooperative bilevel architecture search combining an Illumination Estimation and a Noise Removal Module to optimize enhancement performance.
• The paper demonstrates significant gains in PSNR and SSIM on benchmarks like MIT-Adobe 5K and LOL, outperforming conventional deep learning approaches.

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

The paper presents a novel approach, Retinex-inspired Unrolling with Architecture Search (RUAS), aimed at addressing the low-light image enhancement problem. Traditionally, this task has been tackled using various deep learning models, which often demand extensive architecture engineering and suffer from computational inefficiencies. RUAS proposes an efficient and robust method to counter these challenges by combining principles of Retinex theory with optimization unrolling and cooperative architecture search.

Methodology

RUAS capitalizes on Retinex theory, which separates the observed image into reflectance and illumination components, to model the intrinsic characteristics of underexposed images. The proposed framework involves two main components: the Illumination Estimation Module (IEM) and the Noise Removal Module (NRM), structured via an unrolling strategy.

1. Illumination Estimation Module (IEM): This module estimates the illumination map iteratively, using a warm-start strategy that adjusts the initial estimates and refines them via a parameterized CNN architecture. This approach optimizes based on the intrinsic structure of low-light images, enabling more effective enhancement of illumination than traditional methods.
2. Noise Removal Module (NRM): Tailored for handling noise inherently present in low-light scenarios, the NRM employs a similar unrolling strategy. This module suppresses noise through iterative optimization, again parameterized via a CNN that adapts to various noise characteristics.

Cooperative Prior Architecture Search

A crucial innovation in RUAS is the cooperative bilevel search strategy for architecture discovery. Contrary to existing neural architecture search (NAS) algorithms that depend heavily on data without incorporating domain-specific knowledge, RUAS designs a search space that explicitly models low-light image enhancement tasks. By formulating architecture search as a cooperative optimization problem between IEM and NRM, the framework efficiently discovers architectures that balance illumination estimation and noise removal effectively.

Results and Comparisons

RUAS exhibits superior performance compared to existing state-of-the-art methods on commonly used datasets like MIT-Adobe 5K and LOL. Quantitative results illustrate significant improvements in PSNR and SSIM metrics. Notably, RUAS maintains a competitive edge in scenarios involving varying light intensities and substantial noise, outperforming manual architecture design approaches.

Implications and Future Directions

The contributions of RUAS emphasize the potential of integrating classical image processing theories with modern deep learning techniques. By leveraging principled optimization unrolling with automated architecture discovery, RUAS offers a computationally efficient solution without sacrificing performance quality.

The advancements presented in this paper pave the way for future research in low-level vision tasks. Prospective studies could extend this cooperative search paradigm to other domains within image processing, such as super-resolution or denoising tasks. Furthermore, exploring the transferability of discovered architectures across different imaging conditions and applications could broaden the utility of the RUAS framework in diverse contexts.

Conclusion

This paper introduces a well-structured and efficient approach to low-light image enhancement by synthesizing Retinex-inspired models with cooperative architecture search, surpassing existing methodologies in both performance and computational efficiency. RUAS stands out for its ability to autonomously uncover high-performing architectures tailored to the nuanced challenges of low-light enhancement, marking a meaningful step forward in this domain.