Geometry-aware view reconstruction network for light field image compression.

Area of Science:

• Computer Vision and Image Processing
• Data Compression and Transmission

Background:

• Light Field (LF) imaging captures rich spatial and angular light ray information, enabling advanced applications.
• Existing view reconstruction methods for LF compression face limitations due to occlusion and under-utilization of structural information, impacting reconstruction quality and compression efficiency.

Purpose of the Study:

• To propose a novel geometry-aware view reconstruction network to address the limitations of current LF compression techniques.
• To improve the quality of reconstructed LF views and enhance the overall efficiency of LF compression.

Main Methods:

• A two-stage reconstruction process is introduced: geometry-aware reconstruction and texture refinement.
• The geometry-aware stage employs a multi-stream framework to leverage LF spatial-angular, location, and geometry information from sparsely encoded views.
• The texture refinement stage fuses this rich LF information to further enhance reconstruction quality.

Main Results:

• The proposed method significantly improves LF view reconstruction quality compared to existing approaches.
• Experimental results demonstrate superior rate-distortion performance, indicating higher compression efficiency.
• Enhanced perceptual quality of reconstructed views is achieved while reducing bitrate requirements.

Conclusions:

• The geometry-aware view reconstruction network effectively addresses occlusion and structural information under-exploitation in LF compression.
• The method offers a promising solution for efficient LF storage and transmission by balancing compression rates and reconstruction fidelity.
• This approach advances the state-of-the-art in Light Field compression technology.