Deep Spatial-Angular Regularization for Light Field Imaging, Denoising, and Super-Resolution.

Area of Science:

• Computer Vision
• Image Reconstruction
• Computational Imaging

Background:

• Coded aperture imaging enables 4-D light field (LF) capture by compressing data into 2-D measurements.
• Current reconstruction algorithms for coded aperture LFs suffer from limited quality due to algorithmic bottlenecks.

Purpose of the Study:

• To develop a novel learning-based framework for high-quality 4-D LF reconstruction from coded aperture acquisitions.
• To address the limitations of existing reconstruction algorithms in terms of quality and efficiency.

Main Methods:

• A deep learning framework incorporating measurement observations is proposed, formulating reconstruction as an inverse problem with implicit regularization.
• A deep efficient spatial-angular separable convolutional sub-network is used for regularization, employing local and global residual learning.
• The framework is extended for LF denoising and spatial super-resolution tasks.

Main Results:

• The proposed method significantly outperforms state-of-the-art approaches in both quantitative (PSNR/SSIM) and qualitative assessments.
• Reconstructed LFs exhibit improved fidelity and better preservation of the parallax structure.
• Demonstrated effectiveness on both real and synthetic LF datasets.

Conclusions:

• The developed learning-based framework offers a superior solution for coded aperture 4-D LF reconstruction.
• The approach effectively overcomes the limitations of traditional reconstruction methods.
• The framework's adaptability to denoising and super-resolution highlights its versatility in image processing.