RCA-LF: Dense Light Field Reconstruction Using Residual Channel Attention Networks

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Multi-view image reconstruction is crucial for enhancing application efficiency.
• Existing methods struggle with the angular-spatial resolution discrepancy in high-density light fields.
• Light field (LF) imaging, particularly from Lytro Illum cameras, presents unique reconstruction challenges.

Purpose of the Study:

• To develop an effective solution for reconstructing high-density light field images.
• To address the inherent resolution limitations of captured light field data.
• To improve texture detail restoration in light field reconstruction tasks.

Main Methods:

• Introduction of the Residual Channel Attention Light Field (RCA-LF) structure.
• Grouping view images into stacks to leverage epipolar information.
• Utilizing 2D convolution layers for feature extraction from stacked views.
• Employing a channel attention mechanism to weigh inter-view features and enhance texture.

Main Results:

• The proposed RCA-LF model demonstrates superior performance in light field reconstruction.
• Experimental results show significant improvements in both visual and numerical evaluations.
• The method effectively restores texture details by learning inter-view relationships.

Conclusions:

• The RCA-LF structure provides a robust solution for high-density light field reconstruction.
• The channel attention mechanism is key to improving texture restoration and overall reconstruction quality.
• This work advances the state-of-the-art in light field image processing.