A Spatio-Temporal Recurrent Alignment for Video Super-Resolution

The proliferation of high resolution displays has made the demand for high-quality and computationally appropriate video super-resolution increase. Achieving top-notch video super-resolution boils down to effectively using information between frames. Yet, many current deep learning-based methods depend on complex network structures, such as bidirectional optical flow estimation, modulated deformable convolution, and cross-frame attention-based transformers. This paper introduces a practical video super-resolution method that leverages deep prior restoration information through the learning of spatio-temporal recurrent alignment. The features of our approach enable us to achieve excellent performance without the need for an expensive motion compensation stage. The algorithm's appeal for video super-resolution applications lies in its attractive combination of superior performance and reduced computational load. Our experimental results show the effectiveness and efficiency of video super-resolution, especially in real-world scenarios with diverse degradation types and complex motion compensation requirements.