Sparse Gradient Regularized Deep Retinex Network for Robust Low-Light Image Enhancement.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Existing data-driven methods for low-light image enhancement often produce suboptimal results, including amplified noise, reduced contrast, and color distortion.
• The lack of a clear objective function for low-light enhancement contributes to these undesirable outcomes.

Purpose of the Study:

• To develop an end-to-end network for single-image low-light enhancement guided by signal priors and incorporating layer-specific constraints.
• To address limitations of previous methods by improving noise handling, contrast, and color fidelity.

Main Methods:

• Designed an end-to-end network inspired by Retinex theory, featuring signal prior-guided layer separation and data-driven mapping.
• Introduced a Sparse Gradient Minimization sub-Network (SGM-Net) for extracting illumination maps by preserving edges and minimizing low-amplitude structures.
• Employed Enhance-Net and Restore-Net for predicting enhanced illumination and reflectance maps, respectively, with layer-specified constraints.

Main Results:

• The proposed model effectively enhances contrast in illumination maps and reduces noise in reflectance maps.
• Evaluations on synthetic and real-world low-light images, including those with noise and compression artifacts, demonstrate significant improvements.
• The method achieves superior visual quality and outperforms state-of-the-art low-light enhancement techniques.

Conclusions:

• The developed signal prior-guided network with layer-specified constraints offers an effective solution for single-image low-light enhancement.
• The approach successfully mitigates noise amplification, contrast degradation, and color bias common in previous methods.
• This work advances the field of low-light image enhancement through a novel combination of layer separation and data-driven mapping.