Research On Video Super-resolution Reconstruction Method Based On Deep Recursive Network

How to use the redundant high and low-resolution frame information between frames more effectively in different scenarios is a technical focus of video super-resolution reconstruction.In this thesis,several new video super-resolution methods are designed for scenes with small inter-frame motion variations,blur,large motion and occlusion based on the framework of deep recursive networks.The proposed methods can effectively enhance the video super-resolution reconstruction effect and reconstruct the edges and texture details relatively clearly.The main work is specified as follows:(1)Existing video super-resolution reconstruction methods do not make full use of the large amount of redundant information in the fusion stage and do not consider the balance between reconstruction performance and speed well in scenes such as relatively still and small inter-frame motion changes.In this thesis,a multi-level attention network-based video super-resolution reconstruction method is designed.The method is based on a recursive-based adaptive aggregation network and an adaptive multi-level attention module,resulting in residual space fusion and inter-frame information fusion,which can better extract contextual and inter-channel information and enhanced ability to extract linked information over time.Experimental results on a standard dataset demonstrate that the proposed method can effectively enhance video super-resolution reconstruction and achieve a better balance between reconstruction performance and speed.(2)In the previous study,when scenes with large inter-frame motion variations,distant targets and blurred targets occur,the lack of inter-frame alignment operations can greatly affect the effectiveness of subsequent video super-resolution reconstruction.In this thesis,we design a video super-resolution reconstruction method based on gated high and low-resolution frames.The core idea is to introduce a gating mechanism,while using the information of high and low-resolution neighbouring frames to perform motion compensation in an adaptive manner.Also,this thesis utilises a scaled local hierarchical salient feature fusion network for fusing features of aligned video frames to extract contextual information and locally salient features to improve the reconstruction quality of the video.Compared with existing video super-resolution methods,this method achieves better reconstruction performance and clearer edge and texture details.(3)In the previous study and in existing video super-resolution methods most of them take local alignment operations,which can lead to inaccurate motion estimation when there are large motions,occlusions and scenes with large targets in local areas.In addition,when feature fusion,some existing methods do not sufficiently consider redundant information in space and time,inter-frame variability,etc,resulting in poor reconstruction results.In this thesis,a video super-resolution method based on second-order gated global matching and fusion networks is designed.The method achieves more accurate global motion estimation and compensation through a second-order gated global alignment network,avoiding the problem of large errors.The method also makes use of multiple scales for motion estimation,making better use of the similarity information between different scales.In terms of fusion,the idea of global and local is adopted,and global features are used to enhance fusion,and then refinement of locally significant features is carried out to obtain locally significant high frequency detail information.This method can obtain better spatio-temporal interaction information as well as texture and edge information,and obtain better a reconstruction effect.Compared with the previous study method,the improvement is 0.52 d B and 0.10 d B respectively.