High-order relational generative adversarial network for video super-resolution

Abstract

Video super-resolution can reconstruct a sequence of high-resolution frames with temporally consistent contents from their corresponding low-resolution sequences. The key challenge for this task is how to effectively utilize both inter-frame temporal relations and intra-frame spatial relations. The existing methods for super-resolving the videos commonly estimate optical flows to align the features of multiple frames based on temporal correlations. However, motion estimation is often error-prone and hence largely hinders the recovery of plausible details. Moreover, high-order contextual dependencies in the feature space are rarely exploited for further enhancing the spatio-temporal information fusion. To this end, we propose a novel generative adversarial network to super-resolve low-resolution videos, which makes full use of patch embeddings and is effective in exploring high-order spatio-temporal relations of the feature patches. Specifically, a motion-aware relation module is designed to handle the alignment between neighboring frames and reference ones. Depending on a patch-matching strategy for adaptive selection of multiple most similar patches, the cross-scale graph is constructed to reliably aggregate these patches using a feature pyramid. Based on the structure of multi-scale graph, a context-aware relation module is developed to capture high-order dependencies among resulting warped patches for better leveraging long-range complementary contexts. To further enhance reconstruction ability, the temporal position information of video sequences is also encoded into this module. Dual discriminators with cycle consistent constraints are adopted to provide more informative feedback to the generator while maintaining the global coherence. Extensive experiments have demonstrated the effectiveness of the proposed method in terms of quantitative and qualitative evaluation metrics.