Denoising 3D integral images by a single-shot unsupervised deep neural network

Area of Science:

• Optics and Photonics
• Computer Vision
• Machine Learning

Background:

• Integral imaging captures 3D radiance fields but is prone to noise, degrading image quality.
• Existing deep learning denoising methods require extensive ground-truth data, limiting their applicability.
• Current methods are often restricted to specific imaging conditions encountered during training.

Purpose of the Study:

• To develop an unsupervised deep learning method for integral imaging denoising.
• To overcome the limitations of data scarcity and condition-specific training in current methods.
• To improve the quality of integral images for better 3D feature extraction and visualization.

Main Methods:

• Proposed a novel single-shot unsupervised deep learning technique for integral imaging denoising.
• Utilized a Noise2Noise approach adapted for single integral image acquisition.
• Leveraged the inherent similarity between elemental and integral imaging properties.

Main Results:

• Successfully denoised integral images using a single acquired shot.
• Demonstrated adaptation to specific image acquisition conditions.
• Showcased increased image quality without reliance on clean images or noise models.

Conclusions:

• The proposed unsupervised method effectively denoises integral imaging data.
• This approach addresses the critical challenge of limited ground-truth data in integral imaging.
• The technique offers a flexible and robust solution for enhancing 3D imaging quality across various applications.