Pre-Demosaic Graph-Based Light Field Image Compression.

Area of Science:

• Computer Vision
• Image Processing
• Data Compression

Background:

• Plenoptic light field (LF) cameras capture directional rays using microlenses.
• Conventional LF processing involves demosaicking and conversion to sub-aperture images (SAIs).
• Existing methods often require pre-processing, impacting efficiency.

Purpose of the Study:

• To propose a new LF image coding scheme that bypasses pre-processing.
• To improve coding efficiency and performance, especially at high PSNRs.
• To enable direct coding of raw sensor data.

Main Methods:

• A graph lifting transform (GLT) based coding scheme is proposed.
• Raw sensor data is directly mapped to SAIs, creating sparse color pixels.
• A novel intra-prediction scheme exploits spatial correlation, modeling residuals as a Gaussian Markov Random Field (GMRF).
• Optimal graph edge weights are learned via a graph learning method.

Main Results:

• The proposed method codes raw LF data without pre-processing.
• Demosaicking is performed after decoding at the receiver.
• The scheme significantly outperforms conventional LF coding with HEVC at high PSNRs.

Conclusions:

• The novel GLT-based LF image coding scheme offers superior performance for high-quality applications.
• Direct coding of raw sensor data and receiver-side demosaicking are effective strategies.
• The method is particularly beneficial for archiving and instant storage scenarios.