Abstract: Video super-resolution was a work of great practical value. In view of the lack of high-resolution resources in the ultra-high-definition industry, to efficiently utilize the rich temporal correlation information and spatial information between video sequence frames, a video super-resolution reconstruction algorithm based on multi-scale time-domain 3D convolution was proposed. The algorithm extracted the spatiotemporal features of the input low-resolution video sequence frames through the 3D convolution of different time scales. 3D convolution can simultaneously model space and time, which is more suitable for processing video tasks than 2D convolution. After the adaptive motion compensation of two spatio-temporal features extracted in different scales and time domains, the sub-pixel convolutional layer performed resolution enhancement, which was added to the up-sampled input frame to obtain the final reconstructed high-resolution image. The experimental results on the standard data set show that the algorithm can significantly boost visual effects and objective quality evaluation indicators such as peak signal-to-noise ratio and structural similarity, outperforming algorithms such as FSRCNN and EDSR. 

Key words: video super-resolution, deep learning, 3D convolution, multi-scale time domain features, sub-pixel convolution