Probabilistic-Based Learning for Joint Light Field Image Compression and Enhancement Under Low-Light Conditions

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Light field (LF) imaging offers rich spatial and angular data, valuable in challenging low-light conditions.
• LF data face dual challenges: high redundancy requiring efficient compression and quality degradation from poor illumination.
• Degradation weakens inter-view consistency and visual perception, complicating LF image processing.

Purpose of the Study:

• To develop a unified framework for joint compression and enhancement of LF images under low-light conditions.
• To address the coupled challenges of data redundancy and illumination-induced quality degradation.
• To improve the efficiency and visual quality of LF image compression and enhancement.

Main Methods:

• Proposed Probabilistic-based learning for joint LF image compression and enhancement (PrL-LFCE).
• Introduced a probability-based multi-directional feature coupling module for adaptive structure preservation and redundancy reduction.
• Designed a swin-gated enhancement module using attention-guided gating for noise suppression and salient region highlighting.

Main Results:

• PrL-LFCE achieved significant bitrate savings (at least 34.86%) compared to state-of-the-art methods.
• The method maintained excellent visual quality in low-light LF images.
• Demonstrated strong joint compression and enhancement capabilities, outperforming existing approaches.

Conclusions:

• PrL-LFCE effectively handles uncertainty from illumination degradation and compression loss through probabilistic modeling.
• The framework successfully unifies structure-aware compression and feature enhancement.
• PrL-LFCE represents a significant advancement in processing low-light LF images.