Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DI-Retinex: Digital-Imaging Retinex Theory for Low-Light Image Enhancement

Shangquan Sun, Wenqi Ren, Jingyang Peng, Fenglong Song, Xiaochun Cao

TL;DR

This work extends Retinex theory for LLIE to account for digital-imaging distortions by proposing Digital-Imaging Retinex (DI-Retinex), which introduces a non-zero-mean offset term $\boldsymbol{\beta}$ in a pixel-wise linear enhancement model $\mathbf{I}_h \approx \boldsymbol{\alpha} \mathbf{I}_l + \boldsymbol{\beta}$. It develops a zero-shot enhancement pipeline using a lightweight three-convolutional-layer network to predict a and b in a two-parameter contrast-brightness mapping, guided by a masked reverse degradation loss in Gamma space and a variance suppression loss to curb variance amplification. The method achieves state-of-the-art performance among unsupervised/zero-shot LLIE methods on LOL-v1 and LOL-v2 while maintaining small model size and fast inference, and it improves downstream face-detection in low light. The approach offers a practical, lightweight preprocessing step for real-time imaging and downstream vision tasks in challenging lighting conditions, with clear theoretical grounding in the extended DI-Retinex framework.

Abstract

Many existing methods for low-light image enhancement (LLIE) based on Retinex theory ignore important factors that affect the validity of this theory in digital imaging, such as noise, quantization error, non-linearity, and dynamic range overflow. In this paper, we propose a new expression called Digital-Imaging Retinex theory (DI-Retinex) through theoretical and experimental analysis of Retinex theory in digital imaging. Our new expression includes an offset term in the enhancement model, which allows for pixel-wise brightness contrast adjustment with a non-linear mapping function. In addition, to solve the lowlight enhancement problem in an unsupervised manner, we propose an image-adaptive masked reverse degradation loss in Gamma space. We also design a variance suppression loss for regulating the additional offset term. Extensive experiments show that our proposed method outperforms all existing unsupervised methods in terms of visual quality, model size, and speed. Our algorithm can also assist downstream face detectors in low-light, as it shows the most performance gain after the low-light enhancement compared to other methods.

DI-Retinex: Digital-Imaging Retinex Theory for Low-Light Image Enhancement

TL;DR

This work extends Retinex theory for LLIE to account for digital-imaging distortions by proposing Digital-Imaging Retinex (DI-Retinex), which introduces a non-zero-mean offset term in a pixel-wise linear enhancement model . It develops a zero-shot enhancement pipeline using a lightweight three-convolutional-layer network to predict a and b in a two-parameter contrast-brightness mapping, guided by a masked reverse degradation loss in Gamma space and a variance suppression loss to curb variance amplification. The method achieves state-of-the-art performance among unsupervised/zero-shot LLIE methods on LOL-v1 and LOL-v2 while maintaining small model size and fast inference, and it improves downstream face-detection in low light. The approach offers a practical, lightweight preprocessing step for real-time imaging and downstream vision tasks in challenging lighting conditions, with clear theoretical grounding in the extended DI-Retinex framework.

Abstract

Many existing methods for low-light image enhancement (LLIE) based on Retinex theory ignore important factors that affect the validity of this theory in digital imaging, such as noise, quantization error, non-linearity, and dynamic range overflow. In this paper, we propose a new expression called Digital-Imaging Retinex theory (DI-Retinex) through theoretical and experimental analysis of Retinex theory in digital imaging. Our new expression includes an offset term in the enhancement model, which allows for pixel-wise brightness contrast adjustment with a non-linear mapping function. In addition, to solve the lowlight enhancement problem in an unsupervised manner, we propose an image-adaptive masked reverse degradation loss in Gamma space. We also design a variance suppression loss for regulating the additional offset term. Extensive experiments show that our proposed method outperforms all existing unsupervised methods in terms of visual quality, model size, and speed. Our algorithm can also assist downstream face detectors in low-light, as it shows the most performance gain after the low-light enhancement compared to other methods.
Paper Structure (31 sections, 1 theorem, 47 equations, 19 figures, 6 tables)

This paper contains 31 sections, 1 theorem, 47 equations, 19 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Classic Retinex Theory
  4. Model-based LLIE Methods
  5. Data-Drive LLIE Methods
  6. Analysis of Retinex Theory in Low-Light
  7. Imaging Noise
  8. Non-linearity.
  9. Quantization error
  10. Dynamic Range Overflow
  11. Digital-Imaging Retinex Theory
  12. LLIE Model based on DI-Retinex
  13. Method
  14. Contrast Brightness Adjustment Function
  15. Reverse Degradation Loss
  16. ...and 16 more sections

Key Result

theorem 1

Given an underexposed image $\mathbf{I}_l \in \mathbb{R}^{H\times W}$ and one of its possible corresponding properly-exposed images $\mathbf{I}_h \in \mathbb{R}^{H\times W}$, $\exists \boldsymbol \alpha, \boldsymbol \beta \in \mathbb{R}^{H\times W}$ such that the following relation holds,

Figures (19)

  • Figure 1: The trade-off between performance, and inference time and model size. Our method can achieve the best low-light enhancement performance with the smallest parameter number or at the fastest speed consistently.
  • Figure 2: The classic Retinex theory is targeted for scene radiance directly incident into human eye as shown in \ref{['fig:fig1a']}. Some methods, e.g., li2018structurezhang2019kindling consider the noise as shown in \ref{['fig:fig1b']}. However, many other factors including quantization error, non-linearity due to camera response and dynamic range overflow due to physical limitation, are ignored. Our DI-Retinex theory in \ref{['fig:fig1c']} takes all of them into consideration.
  • Figure 3: The statistical experiments for showing the existence of $\boldsymbol \beta$ with a non-neglegible magnitude and the existence of $\boldsymbol \beta'$ with a small magnitude.
  • Figure 4: A visual comparison of enhancement results on LOL-v1. Please zoom in for better visualization.
  • Figure 5: A visual comparison of enhancement results on LOL-v1. Please zoom in for better visualization.
  • ...and 14 more figures

Theorems & Definitions (2)

  • theorem 1
  • proof