Efficient Video Super-Resolution via Hierarchical Temporal Residual Networks

Super-Resolving (SR) video is more challenging compared with image super-resolution because of the demanding computation time. To enlarge a low-resolution video, temporal relationship among frames must be fully exploited. We can model SR as multi-frame problem and use deep learning methods to estimate spatial information. This paper proposes lighter residual network, based on multi-stage back projection for SR. improve block by adding weights adaptive feature tuning, add global & local connections explore deeper representation. jointly learn spatial-temporal maps using proposed Spatial Convolution Packing scheme an attention mechanism extract information from both domains. Different others, our network input multiple obtain super-resolved simultaneously. then further quality self-ensemble enhancement meet videos different motions distortions. Results much experimental work show that approaches give large improvement over other state-of-the-art methods. Compared recent CNN works, save, up 60% time achieve 0.6 dB PSNR improvement.