DCD-Net: Decoupling-Centric Decomposition Network for Low-Light Image Enhancement.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• Low-light image enhancement is crucial for various applications.
• Existing methods often overlook the critical role of the decomposition network in separating illumination and reflectance.
• Current decomposition networks may not strictly adhere to the Retinex theory.

Purpose of the Study:

• To propose a novel Decoupling-Centric Decomposition network (DCD-Net) for low-light image enhancement.
• To address the limitations of existing methods by focusing on illumination-reflectance decoupling refinement.
• To improve the quality of Retinex decomposition without relying on separate enhancement networks.

Main Methods:

• DCD-Net comprises a preprocessing network using self-supervised learning to remove Retinex-incompatible features.
• The decomposition network features a Dual-Gated Directional Reflectance Module (DGD-RM) and Reflectance-Guided Multi-head Self-Attention (RG-MSA) for reflectance restoration.
• Illumination estimation utilizes Discrete Cosine Transform (DCT) for local-global estimates.

Main Results:

• Achieved PSNR/SSIM scores of 20.87/0.770 on LOL v1 and 21.66/0.864 on MIT datasets.
• Obtained an average NIQE score of 3.420 across five unpaired datasets.
• Demonstrated superior performance compared to state-of-the-art methods through qualitative and quantitative analyses.

Conclusions:

• DCD-Net effectively enhances low-light images by prioritizing robust illumination-reflectance decoupling.
• The proposed modules (DGD-RM, RG-MSA, DCT-based illumination estimation) significantly contribute to the network's performance.
• Ablation studies confirm the effectiveness of individual components within the DCD-Net architecture.