| Video super-resolution technology aims to recover high-resolution video frames from a series of low-resolution video frames.Existing research is typically based on the deep learning method of the convolutional neural network,which uses the temporal information of adjacent frames of the video to help the reconstruction of the target frame.In recent years,thanks to the powerful performance of the Transformer structure,the field of computer vision has made significant progress.Therefore,evaluating the performance of the Transformer architecture on video super-resolution tasks is a new paradigm exploration for processing video temporal-series data.At the same time,there are a large number of videos with unknown degradation methods and complex scene switching in the real world.It is difficult for a model trained with a single degradation method to achieve better results on datasets with unknown degradation methods and complex scenes.Therefore,it is necessary to explore a model with better generalization for real-world scene challenges.The research content and contributions of this paper are as follows:(1)A cross-scenario video super-resolution dataset is constructed.Aiming at the relatively single scene in the current mainstream data set,this paper constructs a cross-scenario video test data set and evaluates the performance of the mainstream benchmark model on the cross-scenario test set.On this basis,this paper explores whether the mainstream model is robust to the interference of irrelevant information in cross-scene videos,and finds the distribution preference of crossscene video datasets.This confirms the existence of cross-scenario challenges facing the real world and provides ideas for solving cross-scenario problems.(2)A video super-resolution algorithm based on kernel estimation is proposed.The current video super-resolution models are all trained under the assumption of a fixed degradation method.Aiming at the problem of poor generalization of the model trained in a single degradation method in real-world scenarios,this paper designs a Texture Transformer model with a multi-degenerate kernel prior information.First,the model estimates the degradation kernel in the real-world lowquality video and uses this as the degradation before endowing the model with cognitive ability.Secondly,the texture Transformer structure is designed to effectively extract the spatiotemporal information in the video sequence through the attention mechanism,which helps the high-quality reconstruction of the video.Extensive qualitative and quantitative results demonstrate the superiority of the model.(3)A video super-resolution algorithm with adaptive scenario transformation is proposed.Currently,there are many complex scenes in real-world videos,which limit the model to accept effective information,resulting in much irrelevant content interfering with the generation of target frames.In this paper,we propose a novel method for video super-resolution adaptive to scene transitions.The method includes motion estimation of optical flow,multi-scale feature aggregation,and adaptive attention mechanism.Based on the design of motion trajectory markers,the model can not only effectively extract spatiotemporal information on similar scenarios,but also has stronger robustness to cross-scenarios.Experimental results show that the proposed model greatly reduces performance degradation in cross-scenario super-resolution compared to other methods. |
