Three-dimensional integral imaging-based image descattering and recovery using physics informed unsupervised...

Area of Science:

• Optics and Computer Vision
• Image Restoration and Denoising

Background:

• Image restoration and denoising are challenging problems in optics and computer vision.
• Developing robust, data-efficient systems for image restoration is an active research area.
• Physics-informed deep learning is gaining interest for scientific problems.

Purpose of the Study:

• To introduce a 3D integral imaging-based physics-informed unsupervised CycleGAN algorithm.
• To achieve underwater image descattering and recovery.
• To incorporate physical models into the loss function for degradation parameter significance.

Main Methods:

• Utilized a physics-informed unsupervised CycleGAN (Generative Adversarial Network) with a forward and backward pass.
• Employed an encoder-decoder architecture taking clean/degraded images and depth maps.
• Incorporated a physical model into the loss function to provide physical significance to degradation parameters.

Main Results:

• The proposed model was assessed using a dataset from underwater experiments with varying turbidity.
• The algorithm successfully recovered original images from degraded ones.
• The approach also modeled the distribution of sampled degraded images.

Conclusions:

• The 3D Integral Imaging approach shows promise for underwater image restoration, especially in turbid and partially occluded environments.
• The proposed physics-informed CycleGAN method offers an effective solution for image descattering and recovery.
• This technique advances the field of optical imaging and computer vision for challenging underwater scenarios.