REPAID: resolution-enhanced plenoptic all-in-focus imaging using deep neural networks.

Area of Science:

• Computational Imaging
• Deep Learning
• Image Fusion

Background:

• Classical 2D imaging suffers from limited depth-of-focus, losing details outside the focal plane.
• All-in-focus imaging fuses multi-focus images to achieve extended depth-of-view but often lacks high spatial and temporal resolution for dynamic scenes.
• Existing methods struggle to provide dynamic all-in-focus imaging at both high spatial and temporal resolutions.

Purpose of the Study:

• To develop a novel method for dynamic all-in-focus imaging with high spatial and temporal resolutions.
• To overcome the limitations of conventional all-in-focus imaging techniques for dynamic targets.
• To introduce REPAID (resolution-enhanced plenoptic all-in-focus imaging) using deep neural networks.

Main Methods:

• REPAID reconstructs multi-focus images from a single-shot plenoptic image.
• Deep neural networks, designed for real scenes without ground truth, are used for upsampling.
• The method generates all-in-focus images with high temporal and spatial resolutions.

Main Results:

• REPAID successfully generates high-quality all-in-focus images for both static and dynamic scenes.
• The method achieves high spatial and temporal resolutions in all-in-focus imaging.
• Experiments demonstrate the effectiveness of REPAID with simple optical setups.

Conclusions:

• REPAID provides a promising solution for high-quality all-in-focus imaging, especially for dynamic targets.
• The deep learning approach enables efficient and effective fusion of multi-focus information.
• This technique is valuable for applications requiring imaging of dynamic targets across a large depth-of-view.