Super-Resolution in Still Images and Videos via Deep Learning
Publisher
Metadata
Show full item recordAbstract
The evolution of multimedia systems and technology in the past decade has enabled production and delivery of visual content in high resolution, and the thirst for achieving higher de nition pictures with more detailed visual characteristics continues. This brings attention to a critical computer vision task for spatial up-sampling of still images and videos called super-resolution. Recent advances in machine learning, and application of deep neural networks, have resulted in major improvements in various computer vision applications. Super-resolution is not an exception, and it is amongst the popular topics that have been a ected signi cantly by the emergence of deep learning. Employing modern machine learning solutions has made it easier to perform super-resolution in both images and videos, and has allowed professionals from di erent elds to upgrade low resolution content to higher resolutions with visually appealing picture delity. In spite of that, there remain many challenges to overcome in adopting deep learning concepts for designing e cient super-resolution models. In this thesis, the current trends in super-resolution, as well as the state of the art are presented. Moreover, several contributions for improving the performance of the deep learning-based super-resolution models are described in detail. The contributions include devising theoretical approaches, as well as proposing design choices that can lead to enhancing the existing art in super-resolution. In particular, an e ective approach for training convolutional networks is proposed, that can result in optimized and quick training of complex models. In addition, speci c deep learning architectures with novel elements are introduced that can provide reduction in the complexity of the existing solutions, and improve the super-resolution models to achieve better picture quality. Furthermore, application of super-resolution for handling compressed content, and its functionality as a compression tool are studied and investigated.
Authors
Toutounchi, FCollections
- Theses [4253]