Structure-Preserving Video Super Resolution with Multi-Scale Convolution

Video super resolution technology refers to the reconstruction of low resolution video into high resolution frames. In recent years, the application of deep learning to super-resolution technology has attracted extensive attention. However, the reconstruction effect of the existing model still has some problems such as double shadow, structural loss, and the solutions of problems are relatively rare. In this paper, we propose a new idea to use gradient extraction branches to guide the reconstruction of high resolution frames in backbone networks. The loss function is improved by combining gradient loss with pixel loss to improve convergence ability. Multi-scale convolution is introduced into the alignment module to enlarge the receptive field and improve the performance of the model to extract large motion features. Experimental results show that the model has good performance on REDS4 and Vid4 data sets.