Low-Light Image Enhancement by Retinex-Based Algorithm Unrolling and Adjustment.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Low-light image enhancement (LIE) is crucial for various applications.
• Retinex theory-based deep learning methods offer interpretability but have limitations.
• Existing methods struggle to integrate traditional insights and optimize adjustment steps.

Purpose of the Study:

• To propose a novel deep-learning framework for low-light image enhancement.
• To address limitations in existing Retinex-based deep learning methods.
• To improve both decomposition and adjustment stages for better results.

Main Methods:

• Developed a decomposition network (DecNet) using algorithm unrolling for integrated prior learning.
• Designed adjustment networks considering both global and local brightness for effectiveness.
• Implemented a self-supervised fine-tuning strategy for automated optimization.

Main Results:

• The proposed framework demonstrates superior performance on benchmark LIE datasets.
• Quantitative and qualitative experiments confirm the approach's effectiveness.
• The method achieves promising results without manual hyperparameter tuning.

Conclusions:

• The novel framework significantly advances low-light image enhancement.
• Integrating traditional and deep learning approaches yields improved decomposition and adjustment.
• The self-supervised strategy offers an efficient and effective fine-tuning solution.