Multi-frame Image Super-resolution Reconstruction Based On CNN

With the advent of the contemporary social information age,the most important and common thing in people's lives is the image.Whether in military use or in daily life,images can be everywhere.Therefore,images have become more and more important as a medium for people to obtain information.However,due to image acquisition equipment,storage,transmission and other conditions and cost control factors,people often get low-resolution images,which will produce less pleasant visual effects.If it is in medical imaging,it may sometimes affect the doctor's judgment on the details of the image data,thus affecting the diagnosis of the patient's condition.The higher the resolution of the image,the more pixels it contains and the better the visuals.Therefore,the super-resolution reconstruction technology of images is particularly important.Video is also widely seen in modern life,occupying a very important part of the image,such as in the field of video surveillance,so video super-resolution technology is also developing.This paper mainly describes an improved RDB-based multi-frame image super-resolution reconstruction algorithm.In this paper,the Y channel in YCbCr of three consecutive frames is used as the input image,and the residual dense block is used to extract the features of the image.The residual dense block can not only extract the local features and global features of the image,but also alleviate the problem of gradient disappearance.In many multi-frame image super-resolution reconstruction algorithms,the intermediate frame image is directly residual with the last mainstream operation,or simply the convolution operation is performed on the intermediate frame image,and then the residual operation is performed with the mainstream operation.In order to solve this problem,this paper proposes a convolution block of the intermediate frame image,which is composed of some convolutional layers,a split layer and a concat layer,which can effectively extract more detailed information of the intermediate frame image,thus improve the PSNR value and SSIM value of the restored image.And it also adds a bicubic interpolation image of the intermediate frame after the last deconvolution layer of the network,which can speed up the training speed of the network.Experiments show that the improved RDB-based multi-frame image super-resolution reconstruction algorithm proposed in this paper can obtain relatively higher PSNR and SSIM values than the previous multi-frame image super-resolution reconstruction algorithms,and there is a certain enhancement in the subjective visual effect of the image,which can be closer to the high-resolution original image,and the network of this paper also improves the training speed and accelerates the convergence of the network.