Research On Video Super-resolution Reconstruction Technology Based On Deep Learning

Super-resolution reconstruction technology can reconstruct high-quality images with higher resolution and more detailed information from low-resolution and low-quality images with blurred details.Therefore,it has always been a research hotspot in the field of computer vision.Single-image super-resolution reconstruction performs reconstruction by exploring the self-similarity within the image,which is limited by the limited information of the image itself,and its reconstruction results are also limited to a certain extent,while video super-resolution reconstruction can use more information from adjacent frames to help reconstruction.However,how to efficiently and fully utilize the temporal and spatial correlation information between adjacent frames for reconstruction is an urgent problem to be solved.At present,in video super-resolution reconstruction methods,most of the spatiotemporal information between adjacent frames is extracted by means of registration or optical flow estimation.This method is highly dependent on the accuracy of registration and optical flow estimation.During high-power super-resolution reconstruction,the details of the reconstruction results are lost more.In view of the above problems,this paper studies the video super-resolution reconstruction method based on deep learning technology,and proposes a video super-resolution reconstruction method without calculating optical flow,which can effectively extract the spatiotemporal features between adjacent frames and improve the accuracy of reconstruction results.The main research contents of this paper are as follows:(1)A multi-scale temporal 3D convolutional super-resolution reconstruction method is proposed,which solves the problem of over-dependence on optical flow estimation in video super-resolution reconstruction method and can complete video super-resolution reconstruction without calculating optical flow.Based on the feature that 3D convolution can simultaneously model space and time,and combined with multiple time scales in video sequence,this method can complete the extraction of spatio-temporal features between video sequence frames,and effectively improve the quality of reconstructed frames by utilizing the time correlation and spatial correlation information between sequence frames.Compared with other methods,the experimental results on open data set show that compared with traditional bicubic interpolation method and some deep learning methods,the proposed method can improve the objective evaluation indexes such as PSNR and SSIM,and can reconstruct more detailed texture information in subjective visual effect.(2)A video super-resolution reconstruction method based on extensive self-attention is proposed,which solves the problem of too much detail information loss in reconstruction with large up-sampling factors,and effectively improves the quality of reconstruction results.The method using the attention mechanism of sequence modeling ability of task characteristics,puts forward a kind of extensive attention from the module,can maintain a higher dimension of feature information and a wide range of mapping,by calculating the different space and channel characteristic information on weights,from many features of target screened out the more critical,and complete video super-resolution reconstruction.Compared with other methods on open data sets,the objective evaluation index of reconstruction results of this method is superior to traditional interpolation algorithm and deep learning algorithm such as VESPCN and SOF-VSR,which proves that this method still has a good effect on reconstruction results of large factors.(3)Apply the super-resolution reconstruction method studied in this paper to real-world projects.In the project "Key Mathematical Problems for Massive Multi-source Remote Sensing Data Processing and Its Industrial Applications",the corresponding functional modules are designed and implemented with the super-resolution reconstruction method studied in this paper as the core.The modules have a visual interface and user interaction,including data acquisition,super-resolution reconstruction,video processing and other functional designs.It has high practical value in many fields such as daily life and ultra-high-definition industry.