Light Field Synthesis by Training Deep Network in the Refocused Image Domain.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Light field imaging captures spatial-angular data, enabling applications like refocusing and augmented reality.
• Limited sensor resolution creates a trade-off between spatial and angular resolution in light fields.
• Existing view synthesis methods often prioritize individual view quality over refocused image quality.

Purpose of the Study:

• To propose a novel loss function, refocused image error (RIE), for light field view synthesis.
• To optimize synthesized light field quality specifically within the refocused image domain.
• To address the neglect of refocused image quality in traditional learning-based view synthesis.

Main Methods:

• Developed a new loss function: refocused image error (RIE).
• Analyzed RIE behavior in the spectral domain.
• Evaluated performance on real (INRIA) and synthetic (HCI) light field datasets.
• Utilized objective metrics: MSE, MAE, PSNR, SSIM, and GMSD.

Main Results:

• The proposed RIE loss function leads to improved light field generation.
• Synthesized light fields demonstrate superior quality in the refocused image domain.
• Experimental results show better refocused images compared to previous methods on standard datasets.

Conclusions:

• Optimizing for refocused image quality is crucial for effective light field synthesis.
• The RIE loss function offers a significant advancement in generating high-quality refocused images from synthesized light fields.
• This approach enhances the practical utility of light field imaging applications.