Efficient, Low-Cost, Real-Time Video Super-Resolution Network

Video Super-Resolution (VSR) task aims to reconstruct missing high-frequency information lost in degradation. Researchers have proposed many excellent models. However, these models require large memory and high computational cost. In this paper, we propose a novel VSR model called StudentVSR (StuVSR) which is a unidirectional recurrent network. To guarantee StuVSR can generate sufficient high-frequency information, we propose Inceptual Attention (IA) mechanism. Meanwhile, to compress the model size, we utilize the idea of knowledge distillation. We take an auto-encoder network as teacher and redesign the knowledge distillation mode. StuVSR employs extremely small parameters and accomplishes the VSR task in a rapid manner. StuVSR can generate 30-frame-per-second (FPS) 1080p-2k videos in real-time. We conduct comparison experiments to prove the superiority of StuVSR and StuVSR achieves the highest Peak Signal to Noise Ratio (PSNR) score among 16 state of the arts. We also explore the function of the inceptual attention and the knowledge distillation mode through ablation experiments. We will publish the codes at https://github.com/Dawn3474/StuVSR.