Skip to Main content Skip to Navigation
New interface

Selected topics in learning-based coding for light field imaging

Abstract : The current trend in imaging technology is to go beyond the 2D representation of the world captured by a conventional camera. Light field technology enables us to capture richer directional cues. With the recent availability of hand-held light field cameras, it is possible to capture a scene from various perspectives with ease at a single exposure time, enabling new applications such as a change of perspective, focusing at different depths in the scene, and editing depth-of-field.Whereas the new imaging model increases frontiers of immersiveness, quality of experience, and digital photography, it generates huge amounts of data demanding significant storage and bandwidth resources. To overcomethese challenges, light fields require the development of efficient coding schemes.In this thesis, we explore deep-learning-based approaches for light field compression. Our hybrid coding scheme combines a learning-based compression approach with a traditional video coding scheme and offers a highly efficient tool for lossy compression of light field images. We employ an auto-encoder-based architecture and an entropy constrained bottleneck to achieve particular operability of the base codec. In addition, an enhancement layer based on a traditional video codec offers fine-grained quality scalability on top of the base layer. The proposed codec achieves better performance compared to state-of-the-art methods; quantitative experiments show, on average, more than $30%$ bitrate reduction compared to JPEG Pleno and HEVC codecs.Moreover, we propose a learning-based lossless light field codec that leverages view synthesis methods to obtain high-quality estimates and an auto-regressive model that builds probability distribution for arithmetic coding. The proposed method outperforms state-of-the-art methods in terms of bitrate while maintaining low computational complexity.Last but not least, we investigate distributed source coding paradigm for light field images. We leverage the high modeling capabilities of deep learning methods at two critical functional blocks in the distributed source coding scheme: for the estimation of Wyner-Ziv views and correlation noise modeling. Our initial study shows that incorporating a deep learning-based view synthesis method into a distributed coding scheme improves coding performance compared to the HEVC Intra. We achieve further gains by integrating the deep-learning-based modeling of the residual signal.
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Friday, November 18, 2022 - 2:39:51 PM
Last modification on : Tuesday, November 22, 2022 - 8:37:50 AM


Version validated by the jury (STAR)


  • HAL Id : tel-03860034, version 1


Milan Stepanov. Selected topics in learning-based coding for light field imaging. Signal and Image processing. Université Paris-Saclay, 2022. English. ⟨NNT : 2022UPASG050⟩. ⟨tel-03860034⟩



Record views


Files downloads