GPRF-HPNet Physics-guided texture-aware fusion for real-world underwater image enhancement

Area of Science:

• Computer Vision
• Image Processing
• Robotics

Background:

• Underwater images suffer from degradation due to wavelength-dependent attenuation and scattering.
• These issues cause color casts, reduced contrast, and veiling effects, impacting robotic applications.
• Existing methods often struggle to balance global color consistency with local texture preservation.

Purpose of the Study:

• To develop an advanced underwater image enhancement technique that addresses color casts and preserves fine texture details.
• To improve the stability and reliability of computer vision modules for underwater robots.
• To propose a novel Physics-Guided Texture-Aware Fusion for Real-World Underwater Image Enhancement (GPRF-HPNet) network.

Main Methods:

• A YCbCr domain attenuation map guides color correction, followed by entropy-driven dual histogram contrast enhancement.
• Gradient-weighted wavelet fusion combines intermediate images, preserving structural information and fine details.
• A parallel fusion network with Gabor texture maps and a base branch extracts joint features for enhanced image reconstruction.

Main Results:

• The proposed GPRF-HPNet method demonstrated consistent performance gains across six metrics, including UIQM and UCIQE.
• Enhanced images showed improved quality in terms of color consistency and texture detail preservation.
• The method achieved more reliable performance in downstream tasks like geometric rotation estimation and edge detection.

Conclusions:

• The GPRF-HPNet method effectively enhances real-world underwater images by balancing global color consistency and local texture details.
• The proposed approach offers robust generalization across diverse underwater scenes.
• This technique significantly improves the performance of underwater robotic vision systems.