Learning-based high-efficiency compression framework for light field videos
Journal Publication ResearchOnline@JCUAbstract
The massive amount of data usage for light field (LF) information poses grand challenges for efficient compression designs. There have been several LF video compression methods focusing on exploring efficient prediction structures reported in the literature. However, the number of possible prediction structures is infinite, and these methods fail to fully exploit the intrinsic geometry between views of an LF video. In this paper, we propose a deep learning-based high-efficiency LF video compression framework by exploiting the inherent geometrical structure of LF videos. The proposed framework is composed of several crucial components, namely sparse coding based on a universal view sampling method (UVSM) and a CNN-based LF view synthesis algorithm (LF-CNN), a high-efficiency adaptive prediction structure (APS), and a synthesized candidate reference (SCR)-based inter-frame prediction strategy. Specifically, instead of encoding all the views in an LF video, only parts of views are compressed while the remaining views are reconstructed from the encoded views with LF-CNN. The prediction structure of the selected views is able to adapt itself to the similarity between views. Inspired by the effectiveness of view synthesis algorithms, synthesized results are served as additional candidate references to further reduce inter-frame redundancies. Experimental results show that the proposed LF video compression framework can achieve an average of over 34% bitrate savings against state-of-the-art LF video compression methods over multiple LF video datasets.
Journal
Multimedia Tools and Applications
Publication Name
N/A
Volume
81
ISBN/ISSN
1573-7721
Edition
N/A
Issue
6
Pages Count
34
Location
N/A
Publisher
Springer
Publisher Url
N/A
Publisher Location
N/A
Publish Date
N/A
Url
N/A
Date
N/A
EISSN
N/A
DOI
10.1007/s11042-022-11955-8