Light Field Reconstruction Using Convolutional Network on EPI and Extended Applications.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Light field reconstruction from sparse views is challenging due to information loss and aliasing.
• Direct upsampling methods often introduce ghosting artifacts, degrading reconstruction quality.

Purpose of the Study:

• To develop a novel convolutional neural network (CNN)-based framework for accurate light field reconstruction.
• To address the information asymmetry and undersampling issues in angular dimensions of epipolar plane images (EPIs).

Main Methods:

• A "blur-restoration-deblur" framework is proposed, leveraging angular restoration on EPIs.
• The method involves extracting low-frequency spatial components via blurring, restoring angular details with a CNN, and recovering spatial high frequencies through deblurring.

Main Results:

• The framework effectively suppresses ghosting artifacts common in angular super-resolution.
• Evaluations on synthetic, real-world, and microscope light field data demonstrate superior performance and robustness compared to state-of-the-art methods.
• Extended applications include depth enhancement, interpolation, and novel rendering for large disparities.

Conclusions:

• The proposed CNN-based framework offers an efficient and effective solution for light field reconstruction from sparse views.
• The "blur-restoration-deblur" approach significantly improves reconstruction quality by mitigating ghosting and restoring details.
• The framework shows promise for advanced applications in computer vision and imaging.