Super-resolution network-based fractional-pixel motion compensation

In video coding, the efficiency of inter-prediction in highly efficient video coding (HEVC) is improved by fractional-pixel motion compensation and is widely used. The traditional fractional-pixel motion compensation method usually interpolates fractional pixels from the integer pixels. However, the existing fractional interpolation methods based on deep learning either only generates half pixels or needs to train corresponding models for corresponding sub-pixel positions. In this paper, we propose to use the super-resolution reconstruction model as the basic structure of the network to solve the problem of fractional-pixel motion compensation in HEVC. This is constructed from a detail component and a super-resolution network to make feature extraction accurate and generate fractional pixels. Firstly, we extract the detail components of adjacent frames as complementary information to achieve accurate feature extraction. Secondly, the data expansion methods are utilized to make full use of the dataset when generating training data. Finally, we set the upscaling factor to four in the deconvolution layer to generate fractional pixels. Experimental results show that this method can save more bits compared with HEVC.