Examination Committee

Prof Huamin QU, CSE/HKUST (Chairperson)
Prof Dit-Yan YEUNG, CSE/HKUST (Thesis Supervisor)
Prof Shaojie SHEN, ECE/HKUST

Abstract

The goal of video super-resolution (VSR) is to predict the high-resolution (HR) video sequence from a low-resolution (LR) video sequence input. In the past several decades, many researchers have used multi-image super-resolution (MISR) methods to do VSR, which have not considered the temporal correlations between the reconstructed HR frames. And traditional MISR methods have the drawback of demanding highly accurate motion estimation, which causes these methods to be unable to well handle complex motion situations. In our work, we use bidirectional Convolutional Long Short Term Memory networks (ConvLSTM) for sequence modelling to do VSR, which can avoid motion estimation and achieve better temporal dependency modelling than the state of the art VSR method, BRCN. To further improve the super-resolution performance, we propose the adaptation framework for making use of test data’s self-information to adapt our generic model to the concrete test data. Experimental results show that our bidirectional ConvLSTM can achieve a better VSR performance in spatiotemporal domain than the state of the art VSR method, BRCN, and the adaptation framework can enhance previous VSR results further.