Gated Recurrent Networks for Video Super Resolution

Despite the success of Recurrent Neural Networks in tasks involving temporal video processing, few works in Video Super-Resolution (VSR) have employed them. In this work we propose a new Gated Recurrent Convolutional Neural Network for VSR adapting some of the key components of a Gated Recurrent Unit. Our model employs a deformable attention module to align the features calculated at the previous time step with the ones in the current step and then uses a gated operation to combine them. This allows our model to effectively reuse previously calculated features and exploit longer temporal relationships between frames without the need of explicit motion compensation. The experimental validation shows that our approach outperforms current VSR learning based models in terms of perceptual quality and temporal consistency.