Model and data driven approaches to wireless image transmission
File(s)
Author(s)
Karamanli, Can
Type
Thesis
Abstract
This dissertation introduces novel wireless image/video transmission schemes designed to optimize end-to-end reconstruction quality under limited available resources such as memory, bandwidth and power, over static, time-varying, as well as multi-user channels. The proposed approaches address the challenges associated with conventional wireless image transmission by integrating traditional techniques with deep neural network architectures. One of the contributions is the development of ZZCast, a robust low-complexity solution for wireless image transmission. By leveraging the strengths of both traditional techniques and state-of-the-art deep learning methods, ZZCast achieves highly effective image transmission. Extensive experiments have been conducted to validate its effectiveness, and the results demonstrate its promising performance. The rest of the thesis focuses on neural network aided joint source-channel coding schemes combined with model-driven design components. In particular, DeepJSCC-MMSE is proposed for transmitting images over a time-varying fading channel when channel state information (CSI) is available only at the receiver. Model-driven channel equalization is employed together with neural network based DeepJSCC scheme for reduced complexity, and is shown to provide improved performance compared to advance neural network architectures employing channel attention modules. Next, we consider a variable-length DeepJSCC scheme which provides adaptation to the source sample and channel conditions. We also study JSCC over a multiple access channel, and propose a model-driven non-orthogonal multiple access scheme. Through numerical experiments, we show that the proposed approach outperform time-division approach by providing a learned access methodology. Furthermore, the thesis explores advanced techniques including variational autoencoder and diffusion models within the context of wireless image transmission. Through experimentation, the potential of these methods has been demonstrated, showcasing the ability to enhance the performance and capabilities of wireless image transmission systems. Finally, we develop vision transformer(ViT)-based architectures specifically tailored for image transmission tasks. By leveraging the strengths of ViT, we further enhance the efficiency and effectiveness of image transmission systems.
Overall, we believe that this thesis contributes to the fast developing literature on neural network aided JSCC design, particularly focusing on image transmission applications. We hope that our results motivate the practical value of this novel approach, highlight the potential of integrating deep learning architectures with traditional techniques to revolutionize conventional wireless image transmission. The proposed approaches offer promising performance improvements, and pave the way for further advancements in the field. Continued development and refinement of these methodologies, as well as their exploration in various application domains, hold great promise for the future of wireless image transmission.
Overall, we believe that this thesis contributes to the fast developing literature on neural network aided JSCC design, particularly focusing on image transmission applications. We hope that our results motivate the practical value of this novel approach, highlight the potential of integrating deep learning architectures with traditional techniques to revolutionize conventional wireless image transmission. The proposed approaches offer promising performance improvements, and pave the way for further advancements in the field. Continued development and refinement of these methodologies, as well as their exploration in various application domains, hold great promise for the future of wireless image transmission.
Version
Open Access
Date Issued
2023-06-07
Date Awarded
01/11/2023
License URL
Advisor
Gunduz, Deniz
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)