PH-Mamba: Enhancing Mamba With Position Encoding and Harmonized Attention for Image Deraining and Beyond.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Mamba models excel at long-range dependencies with linear complexity.
• Mamba's 1D sequential processing hinders spatial dependency recovery in vision tasks, causing inefficiency.

Purpose of the Study:

• To introduce PH-Mamba, a framework that improves Mamba for image deraining by optimizing spatial dependency modeling.
• To enhance efficiency and accuracy in vision tasks by addressing Mamba's limitations.

Main Methods:

• PH-Mamba integrates position encoding and harmonized attention for image deraining.
• A position-guided hybrid Mamba module (PHMM) encodes spatial coordinates and perturbation features.
• A harmonized Transformer within PHMM focuses on uncertain regions and suppresses noise.

Main Results:

• PH-Mamba achieves superior performance in low-level image restoration benchmarks.
• Compared to NeRD, PH-Mamba shows a 0.60 dB PSNR improvement.
• PH-Mamba significantly reduces parameters (88.9%), computation (36.2%), and inference time (63.0%).

Conclusions:

• PH-Mamba offers a more efficient and accurate approach to image deraining and other vision tasks.
• The framework's combination of global positional guidance and local features strengthens contextual learning.
• PH-Mamba represents a significant advancement over existing Mamba-based methods for vision applications.