Mueller transform matrix neural network for underwater polarimetric dehazing imaging

Area of Science:

• Optics
• Image Processing
• Computer Vision

Background:

• Polarization dehazing imaging is crucial for restoring images in scattering media like turbid water.
• Existing learning-based methods often lack physical principles, limiting underwater polarimetric dehazing performance.

Purpose of the Study:

• To propose a novel Mueller Transform Matrix Network (MTM-Net) for improved underwater polarimetric image recovery.
• To integrate physical dehazing models into a deep learning framework for enhanced performance.

Main Methods:

• Developed a Mueller Transform Matrix Network (MTM-Net) incorporating the Mueller matrix method for physical dehazing.
• Employed a combined content and pixel loss function for detailed image recovery.
• Utilized inverse residuals and channel attention for accelerated network performance.

Main Results:

• The MTM-Net demonstrated significantly improved recovery performance compared to existing methods.
• Ablation experiments and comparative tests validated the network's effectiveness.
• The method achieved better recovery effects with preserved image quality.

Conclusions:

• The proposed MTM-Net offers a physics-informed approach to underwater polarimetric dehazing.
• This method enhances image recovery in turbid water environments.
• The study expands the capabilities of polarimetric dehazing techniques.