Video Super-Resolution with Convolutional Neural Networks
Event details
| Date | 24.11.2016 |
| Hour | 14:00 › 16:00 |
| Speaker | Ruofan Zhou |
| Location | |
| Category | Conferences - Seminars |
EDIC Candidacy Exam
Exam President: Prof. Pascal Fua
Thesis Director: Prof. Sabine Süsstrunk
Co-examiner: Prof. Wenzel Jakob
Background papers
A Bayesian Approach to Adaptive Video Super Resolution,by Liu C., Sun, D.
Super-resolution: a comprehensive survey,by Nasrollahi K., Moeslund, T.
Video Super-Resolution with Convolutional Neural Networks, by Kappeler, A., et al.
Abstract
Super-resolving real-world video sequences remains to be a challenging research task in the field of computer vision. Many of the early works focus on model-based methods such as maximum a posteriori and adaptive filtering that can achieve promising results but highly rely on the assumption on the data. As nowadays data-driven method convolutional neural networks (CNN) have so far been successfully applied to image super-resolution (SR) as well as other restoration tasks, we propose a CNN solution for video super-resolution. In this proposal, we firstly present a conventional adaptive video super-resolution algorithm which uses a Bayesian approach [1]. Secondly, we explore the early works of super-resolution [2] research field comprehensively and discuss several key factors such as motion estimation, data statistics and evaluation metrics. Finally, we review a recent work [3] that introduce CNN in part of their video super-resolution framework. Based on the ideas from these three papers and the results from our current work, we conclude with a discussion of potential further directions.
Exam President: Prof. Pascal Fua
Thesis Director: Prof. Sabine Süsstrunk
Co-examiner: Prof. Wenzel Jakob
Background papers
A Bayesian Approach to Adaptive Video Super Resolution,by Liu C., Sun, D.
Super-resolution: a comprehensive survey,by Nasrollahi K., Moeslund, T.
Video Super-Resolution with Convolutional Neural Networks, by Kappeler, A., et al.
Abstract
Super-resolving real-world video sequences remains to be a challenging research task in the field of computer vision. Many of the early works focus on model-based methods such as maximum a posteriori and adaptive filtering that can achieve promising results but highly rely on the assumption on the data. As nowadays data-driven method convolutional neural networks (CNN) have so far been successfully applied to image super-resolution (SR) as well as other restoration tasks, we propose a CNN solution for video super-resolution. In this proposal, we firstly present a conventional adaptive video super-resolution algorithm which uses a Bayesian approach [1]. Secondly, we explore the early works of super-resolution [2] research field comprehensively and discuss several key factors such as motion estimation, data statistics and evaluation metrics. Finally, we review a recent work [3] that introduce CNN in part of their video super-resolution framework. Based on the ideas from these three papers and the results from our current work, we conclude with a discussion of potential further directions.
Practical information
- General public
- Free
Contact
- Cecilia Chapuis EDIC