Interpretable Optimization-Inspired Unfolding Network for Low-Light Image Enhancement.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Retinex-based methods are effective for low-light image enhancement (LLIE) but lack adaptivity and efficiency due to hand-crafted priors and conventional optimization.
• Existing methods struggle with adaptive layer decomposition for optimal LLIE.

Purpose of the Study:

• To propose an adaptive Retinex-based deep unfolding network (URetinex-Net++) for efficient and effective low-light image enhancement.
• To improve upon previous URetinex-Net by addressing color defects and enhancing overall performance.

Main Methods:

• Developed URetinex-Net++, a deep unfolding network that decomposes low-light images into reflectance and illumination layers using implicit priors.
• Introduced three learning-based modules for data-dependent initialization, efficient unfolding optimization, and component adjustment.
• Incorporated a cross-stage fusion block to mitigate color defects.

Main Results:

• The proposed unfolding optimization module adaptively fits implicit priors, enabling noise suppression and detail preservation.
• URetinex-Net++ demonstrates superior performance in both visual quality and quantitative metrics compared to state-of-the-art methods.
• The enhanced network achieves boosted performance with minimal parameters and low computational cost.

Conclusions:

• URetinex-Net++ offers an effective and efficient solution for low-light image enhancement by leveraging adaptive deep unfolding.
• The method significantly improves image quality and detail preservation in challenging low-light conditions.
• Extensive experiments validate the superiority of URetinex-Net++ over existing LLIE techniques.