Fast Super-Resolution Algorithm for Real-Time Communication

Video super-resolution aims to restore a high-resolution video frame from multiple low-resolution frames which can effectively improve the perceived quality of the video and enhance the user's visual experience in real-time video communication. Current video super-resolution algorithms pay more attention to super-resolution performance rather than the inference speed. Most of them adopt computationally expensive alignment and fusion module, which leads to high inference time cost and hinders the real-world deployment. Therefore, it is necessary to achieve a balance between inference speed and super-resolution performance. In this paper, we propose a fast video super-resolution network which is achieved through three lightweight alignment methods and implement it on the video restoration algorithm with enhanced deformable convolutional networks (EDVR). We trained the model through the Vimeo-90K training dataset, and tested the algorithm through the Vid4 and Vimeo-90K-T test datasets. The experimental results show that the inference time of the network with our alignment methods can be nearly 38% shorter than original EDVR.