Research on video super-resolution reconstruction based on residual structure encoding and decoding
In this paper,a residual-based codec network REDVSR is proposed to solve the problem of large motion reconstruc-tion and insufficient utilization of long sequence information in video super resolution(VSR).The network is improved on the basis of BasicVSR and is divided into two stages:encoding and decoding.In the coding phase,recurrent neural network,non-local residual neural network block and optical flow network are used to extract aligned low-resolution frame features.In the decoding phase,bidirectional features are integrated,spatio-temporal information is extracted through spatio-temporal attention network,and high-resolution frames are generated by upsampling.Experimental results show that the proposed method achieves higher recon-struction accuracy on public data sets,and outperforms the existing methods in PSNR and SSIM.
video super resolution reconstructionneural networknon-local residual networkspace-time attention mechanism
万方数据
