HRMamba: A Hybrid Retinex and State-Space Model for Underwater Image Enhancement
Summary
This study introduces HRMamba, an efficient deep learning framework for underwater image enhancement. It significantly improves image quality, color fidelity, and visibility while reducing computational load for underwater vision tasks.
Area of Science:
- Computer Vision
- Image Processing
- Deep Learning
Background:
- Underwater image degradation due to light absorption and scattering.
- Limitations of existing deep learning methods: high computational complexity and poor global dependency modeling.
- Need for efficient underwater image enhancement (UIE) in resource-limited environments.
Purpose of the Study:
- To propose a novel hybrid framework, HRMamba, for efficient underwater image enhancement.
- To address the constraints of computational complexity and global dependency modeling in current UIE methods.
- To improve color fidelity, visibility, and overall image quality for underwater vision tasks.
Main Methods:
- Integration of Retinex theory and state-space models (SSMs) in a hybrid framework.
- Utilizing linear-complexity state-space operations for computational efficiency and global dependency modeling.
- Introducing an Illumination Feature Fusion Module (IFFM) for comprehensive feature fusion and an Illumination-Guided Denoising Module (IGDM) for noise suppression.
Main Results:
- HRMamba achieves state-of-the-art enhancement quality with an efficient architecture.
- Demonstrates significant improvements in color fidelity and visibility restoration.
- Substantially reduces computational demands compared to existing methods.
Conclusions:
- HRMamba offers an effective and efficient solution for underwater image enhancement.
- The proposed framework overcomes limitations of previous deep learning approaches.
- HRMamba shows strong potential for practical deployment in underwater applications.
