DRN-VideoSR: a deep recursive network for video super-resolution based on a deformable convolution shared-assignment network.

Video super-resolution(videoSR) usually involves several steps: motion estimation, motion compensation, fusion, and upsampling. Here, we propose a novel architecture for video SR. First, in place of motion estimation and compensation, this architecture is based on a specially designed deformable convolution shared-assignment network. The model does not require warp operation and uses a three-layer pyramid deformable convolution network. Second, inspired by the idea of back-projection and Encoder-Decoder structure, we propose a deep recursive fusion network that fuses multi-frame information for the target frame. The fusion network adopts a Decoder-Encoder structure with shared weights to construct the back-projection network, and concatenates the output of each back-projection layer. This design not only reduces the network requirements, but also deepens the network structure so that it can extract deeper image features and achieve fusion. Extensive evaluations and comparisons with previous methods validate the strengths of this approach and demonstrate that the proposed framework is able to significantly outperform the current state of the art. [ABSTRACT FROM AUTHOR]