Physics informed image restoration under low illumination with simultaneous parameter estimation using 3D integral...

Area of Science:

• Computer Vision
• Optical Science
• Imaging Science

Background:

• Image restoration is crucial for recovering clean images from degraded inputs.
• Traditional and deep learning methods face challenges with complex imaging environments and large dataset requirements.
• Physics-informed approaches offer enhanced performance and uncertainty quantification.

Purpose of the Study:

• To propose a novel physics-informed deep learning approach for image restoration with simultaneous parameter estimation.
• To leverage 3D integral imaging and Bayesian neural networks (BNN) for improved image recovery.
• To address limitations of purely data-driven deep learning methods in physical imaging problems.

Main Methods:

• Developed a physics-informed deep learning framework combining an image-image mapping architecture with a Bayesian neural network (BNN).
• Utilized simulated data based on a physical model for network training.
• Employed 3D integral imaging for simultaneous image restoration and parameter estimation.

Main Results:

• The proposed approach demonstrated promising experimental results in restoring images degraded by low illumination and partial occlusion.
• Achieved significant improvements compared to traditional 2D imaging-based approaches, even with simulated training data.
• Successfully estimated degradation parameters, showcasing the method's utility beyond image restoration.

Conclusions:

• Physics-informed deep learning with simultaneous parameter estimation offers a robust solution for challenging image restoration tasks.
• The method effectively handles degradations and outperforms 2D imaging techniques.
• This approach enhances the practical applicability of deep learning in physical imaging scenarios.