Geometry-Aware Graph Transforms for Light Field Compact Representation.

Area of Science:

• Computer Vision
• Image Processing
• Signal Processing

Background:

• Dense 4D light fields present challenges for energy compaction.
• Existing methods struggle with computational complexity and preserving correlations.

Purpose of the Study:

• To design geometry-aware local graph-based transforms for improved energy compaction in 4D light fields.
• To address computational costs and maintain angular correlation.

Main Methods:

• Constructing local graphs on super-rays for spatially and geometrically correlated pixels.
• Developing separable transforms to overcome computational expense.
• Proposing a novel transform optimization to preserve angular correlation with non-isometric super-pixels.

Main Results:

• The proposed transforms demonstrate significant benefits in energy compaction.
• Experimental results show improved rate-distortion performance compared to state-of-the-art encoders.
• The optimization method effectively preserves angular correlation.

Conclusions:

• The geometry-aware local graph-based transforms offer an effective solution for 4D light field energy compaction.
• The novel optimization technique enhances compression efficiency, particularly for non-isometric super-pixels.
• This work advances the field of light field compression and coding.