Directionally factorized light field reconstruction with cross-epipolar and spatial modeling.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• High-angular-resolution light fields (LFs) are crucial for applications like refocusing, depth estimation, and virtual reality.
• Acquiring densely sampled LFs is hindered by hardware constraints and high costs.

Purpose of the Study:

• To introduce DFCNet, an end-to-end framework for light field angular super-resolution.
• To reconstruct high-resolution LFs from sparse input data efficiently and accurately.

Main Methods:

• Utilizes a hierarchical Residual-in-Residual architecture with Directionally Factorized Convolutional Blocks.
• Incorporates Squeeze-and-Excitation blocks for adaptive feature recalibration.
• Employs a hybrid interpolation-reconstruction strategy for stable angular upsampling.

Main Results:

• DFCNet outperforms state-of-the-art methods in PSNR and SSIM on synthetic and real-world datasets.
• Accurately reconstructs fine textures, occluded regions, and angular structures.
• Demonstrates effective generalization for synthesizing novel views with high fidelity.

Conclusions:

• DFCNet provides a robust and efficient solution for high-quality light field reconstruction.
• The proposed method addresses the challenges of LF acquisition and enhances imaging capabilities.