Light Field Super-Resolution Using a Low-Rank Prior and Deep Convolutional Neural Networks

Area of Science:

• Computer Vision
• Image Processing

Background:

• Light field imaging is gaining traction for computer vision applications.
• High-resolution light field capture faces challenges with spatial and angular resolution trade-offs.

Purpose of the Study:

• To develop a learning-based method for spatial light field super-resolution.
• To restore high-resolution light fields with consistent angular views.

Main Methods:

• Optical flow for light field alignment.
• Low-rank approximation to reduce angular dimension.
• Deep convolutional neural network (DCNN) for embedding restoration.
• Inverse warping and light field inpainting for view reconstruction.

Main Results:

• The proposed method achieves superior performance compared to existing light field super-resolution algorithms.
• A PSNR gain of 0.23 dB over the second-best method was observed.
• Iterative back-projection further enhanced the results as a post-processing step.

Conclusions:

• The learning-based approach effectively addresses the spatial resolution limitations in light field imaging.
• The method ensures consistency across all angular views in the super-resolved light field.
• This work advances the capabilities of light field super-resolution for practical applications.