Self-Supervised Unfolding Network With Shared Reflectance Learning for Low-Light Image Enhancement

Area of Science:

• Computer Vision
• Image Processing

Background:

• Low-light image enhancement (LIE) is crucial for various applications.
• Existing methods often ignore Retinex theory's physical priors or require paired data.

Purpose of the Study:

• To propose a novel self-supervised unfolding network (S2UNet) for low-light image enhancement.
• To address limitations of existing methods, including data dependency and physical prior modeling.

Main Methods:

• Developed a self-supervised unfolding network (S2UNet) based on Retinex theory.
• Formulated a novel optimization model enforcing content consistency under varying illumination.
• Employed gamma correction to create illumination-different image pairs for self-supervision.
• Integrated a self-supervised denoising mechanism to mitigate noise amplification.

Main Results:

• S2UNet demonstrates superior performance over state-of-the-art unsupervised methods.
• Achieved competitive results compared to supervised methods in quantitative metrics and visual quality.
• Extensive experiments on nine benchmark datasets validate the proposed approach.

Conclusions:

• The proposed S2UNet effectively enhances low-light images using self-supervised learning.
• The method successfully models physical priors and reduces reliance on paired data.
• S2UNet offers a robust solution for low-light image enhancement with improved noise suppression.