Research On Application Of Convolutional Neural Network In Video Super Resolution

High-resolution display devices have flourished in recent years,such as 4K(4096 2160),8K(7680 4320)and 10K(10240 4320)TVs,2K(2560 1440)and 4K mobile phones have emerged in daily life one after another.But due to the lack of high-resolution video resources and the heavy burden of storing and transmitting highresolution video on current systems,people are not always able to get high-definition video.In order to fully enhance the visual experience on high-resolution display devices,the research of video super-resolution algorithms is very important.Super-resolution reconstruction technology is to raise single or multiple lowresolution images from the same scene to a high resolution through a certain algorithm,so that the resulting image can have a higher pixel density,more detailed information and more delicate quality.This technology can not only overcome the limitations of the hardware system,expand the scope of the engineering application,but also greatly reduce the cost,it is worth studying and exploring in depth.When super-resolution algorithm is applied to video,it will not only utilize the information of the current frame,but also the redundant information between adjacent frames based on the characteristics of the video to recover the high frequency part of the current frame.Video super-resolution tasks are inherently more difficult due to additional motion estimation and multi-frame information fusion issues,so it is not widely researched as single super-resolution.The research work of this paper mainly focuses on the above two problems and studies the application of convolutional neural network as an emerging technology in video super-resolution.The research work in this paper mainly includes:1.The existing single image super-resolution model SRCNN based on the convo-lution neural network and its improved accelerated version FSRCNN are analyzed in detail.Based on this,combined with the characteristics of video images,three video superresolution models based on CNN are proposed and compared.2.Motion compensation becomes difficult when large motion and motion blur are present in the video,super-resolution reconstructed images may therefore have some undesirable boundary effects or artificial artifacts.To solve this problem,an adaptive motion compensation strategy is introduced to reduce the effect of mis-registration between adjacent frames.3.The trained single image super-resolution parameter model is modified and then applied to the initialization of the video super-resolution model,further improving the training efficiency and accuracy of the model.4.The CNN-based optical flow estimation method is used in the motion compen-sation part of the proposed video super-resolution model,and an improved acceleration model is proposed to further improve the execution efficiency of the model.Through the comparison and analysis in the experiment part of this article,we can see that the video super-resolution model proposed in this paper has a good performance in terms of subjective evaluation and objective evaluation.The further improved acceleration model can also greatly reduce the running time without sacrificing the quality of video super-resolution images,which also provides hope and possibility for real-time super-resolution of video images.