DWMamba: a structure-aware adaptive state space network for image quality improvement.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Deep learning excels at image quality enhancement but faces computational cost and diverse degradation challenges.
• Existing methods struggle with inconsistent color channel and spatial attenuation in degraded images.
• Resource-constrained environments require efficient yet effective image enhancement solutions.

Purpose of the Study:

• To introduce DWMamba, a degradation-aware and weight-efficient Mamba network for robust image quality enhancement.
• To address limitations of current deep learning models in handling complex degradations and computational demands.
• To improve scene understanding in challenging imaging scenarios through advanced image restoration.

Main Methods:

• Developed DWMamba, a Mamba network incorporating an Adaptive State Space Module (ASSM) with dual-stream channel monitoring and soft fusion.
• Implemented ASSM with linear computational complexity to manage non-uniform degradations effectively.
• Introduced a Structure-guided Residual Fusion (SGRF) module using edge priors and region partitioning for selective feature fusion.

Main Results:

• DWMamba demonstrates superior qualitative and quantitative performance in image quality enhancement.
• The network exhibits strong generalization capabilities across diverse extreme lighting conditions.
• Achieved effective restoration of degraded details and enhancement of low-light textures.

Conclusions:

• DWMamba offers a weight-efficient and degradation-aware solution for image quality enhancement.
• The proposed ASSM and SGRF modules effectively address non-uniform degradations and improve feature fusion.
• DWMamba provides a promising approach for accurate scene understanding in challenging imaging environments.