A Retinex-based network for image enhancement in low-light environments

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Existing low-light image enhancement techniques often struggle with detail loss, color distortion, and noise.
• Developing robust methods for improving image quality in low-light conditions remains a significant challenge.

Purpose of the Study:

• To propose a novel neural network-based approach for effective low-light image enhancement.
• To address the limitations of current methods, specifically detail loss, color distortion, and noise.

Main Methods:

• A three-part neural network architecture: decomposition, reflection denoising, and illumination enhancement.
• Utilized Unet3+ with CA attention for denoising reflection components and adaptive mapping curves for illumination enhancement.
• Employed Retinex theory for fusing processed illumination and reflection images.

Main Results:

• The proposed network achieved excellent visual effects in subjective evaluations.
• Demonstrated significant improvements in objective metrics like Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Natural Image Quality Evaluator (NIQE).
• Outperformed several existing methods on public datasets.

Conclusions:

• The proposed neural network effectively enhances low-light images by mitigating common issues like noise and detail loss.
• The method offers a promising solution for improving image quality in challenging lighting conditions.
• Validated through both subjective and objective assessments, showcasing superior performance.