Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Retinexmamba: Retinex-based Mamba for Low-light Image Enhancement

Jiesong Bai, Yuhao Yin, Qiyuan He, Yuanxian Li, Xiaofeng Zhang

TL;DR

This paper tackles the challenge of enhancing low-light images by integrating Retinex theory with a Mamba-based state-space backbone. It introduces RetinexMamba, comprising an Illumination Estimator and a Damage Restorer built on Illumination Fusion State Space Models (IFSSM) and guided by a 2D Selective Scan backbone, with a Fused-Attention mechanism to improve interpretability. Empirical results on the LOL dataset show state-of-the-art quantitative and qualitative gains over Retinex-based deep learning methods, validated by ablation studies. The work promises faster, more interpretable low-light enhancement with potential for practical deployment, while recognizing the need to reduce parameter counts in future work.

Abstract

In the field of low-light image enhancement, both traditional Retinex methods and advanced deep learning techniques such as Retinexformer have shown distinct advantages and limitations. Traditional Retinex methods, designed to mimic the human eye's perception of brightness and color, decompose images into illumination and reflection components but struggle with noise management and detail preservation under low light conditions. Retinexformer enhances illumination estimation through traditional self-attention mechanisms, but faces challenges with insufficient interpretability and suboptimal enhancement effects. To overcome these limitations, this paper introduces the RetinexMamba architecture. RetinexMamba not only captures the physical intuitiveness of traditional Retinex methods but also integrates the deep learning framework of Retinexformer, leveraging the computational efficiency of State Space Models (SSMs) to enhance processing speed. This architecture features innovative illumination estimators and damage restorer mechanisms that maintain image quality during enhancement. Moreover, RetinexMamba replaces the IG-MSA (Illumination-Guided Multi-Head Attention) in Retinexformer with a Fused-Attention mechanism, improving the model's interpretability. Experimental evaluations on the LOL dataset show that RetinexMamba outperforms existing deep learning approaches based on Retinex theory in both quantitative and qualitative metrics, confirming its effectiveness and superiority in enhancing low-light images.

Retinexmamba: Retinex-based Mamba for Low-light Image Enhancement

TL;DR

This paper tackles the challenge of enhancing low-light images by integrating Retinex theory with a Mamba-based state-space backbone. It introduces RetinexMamba, comprising an Illumination Estimator and a Damage Restorer built on Illumination Fusion State Space Models (IFSSM) and guided by a 2D Selective Scan backbone, with a Fused-Attention mechanism to improve interpretability. Empirical results on the LOL dataset show state-of-the-art quantitative and qualitative gains over Retinex-based deep learning methods, validated by ablation studies. The work promises faster, more interpretable low-light enhancement with potential for practical deployment, while recognizing the need to reduce parameter counts in future work.

Abstract

In the field of low-light image enhancement, both traditional Retinex methods and advanced deep learning techniques such as Retinexformer have shown distinct advantages and limitations. Traditional Retinex methods, designed to mimic the human eye's perception of brightness and color, decompose images into illumination and reflection components but struggle with noise management and detail preservation under low light conditions. Retinexformer enhances illumination estimation through traditional self-attention mechanisms, but faces challenges with insufficient interpretability and suboptimal enhancement effects. To overcome these limitations, this paper introduces the RetinexMamba architecture. RetinexMamba not only captures the physical intuitiveness of traditional Retinex methods but also integrates the deep learning framework of Retinexformer, leveraging the computational efficiency of State Space Models (SSMs) to enhance processing speed. This architecture features innovative illumination estimators and damage restorer mechanisms that maintain image quality during enhancement. Moreover, RetinexMamba replaces the IG-MSA (Illumination-Guided Multi-Head Attention) in Retinexformer with a Fused-Attention mechanism, improving the model's interpretability. Experimental evaluations on the LOL dataset show that RetinexMamba outperforms existing deep learning approaches based on Retinex theory in both quantitative and qualitative metrics, confirming its effectiveness and superiority in enhancing low-light images.
Paper Structure (23 sections, 8 equations, 7 figures, 2 tables)

This paper contains 23 sections, 8 equations, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Low-light Image Enhancement
  4. Distribution Mapping Method.
  5. Traditional Model Method.
  6. Deep Learning Method.
  7. State Space models
  8. Method
  9. Retinex-based Framework
  10. Illumination Estimator.
  11. Illumination-Fused Vison Mamba.
  12. Illumination-Fused State Space Models.
  13. Illumination-Fused Attention.
  14. 2D-Selective Scan.
  15. Experiment
  16. ...and 8 more sections

Figures (7)

  • Figure 1: The image above shows a visual comparison between the attention mechanism and 2D selective scanning in Mamba. In Mamba's 2D selective scanning, scanning starts simultaneously from all sides of the image, whereas the attention mechanism calculates attention scores separately from the target view to the global view.
  • Figure 2: Our structural framework is displayed in the two main areas above. It comprises an illumination estimator (a) and a Damage Restorer based on the Illumination Fusion Visual Mamba (IFVM) (b).
  • Figure 3: Our Illumination Fusion State Space Model (IFSSM) integrates lighting features and input vector $x$ using a fused-block, and utilizes the linear 2D Selective Scanning model (SS2D) for feature extraction. In IFA, we treat lighting features as $Q$, and input vectors as $KV$ to calculate attention scores.
  • Figure 4: 3D image display of the HSV color space.
  • Figure 5: The above are the qualitative experimental results on LOLv1. Our method effectively reduced color distortion and enhanced the lighting effects.
  • ...and 2 more figures