Research On Image Super-resolution Reconstruction Based On Deep Learning

With the development of information technology,people have many ways to obtain external information.As an important carrier of information transmission,images have become increasingly important in people's lives.However,due to the limitations of hardware conditions in imaging equipment and the interference of signals during image transmission,the resolution of images obtained in practical applications is often low,and it is difficult to meet people's needs in visual effects.Therefore,if the high-frequency information can be recovered from the low-resolution image,it will help to improve the resolution of the image and make the image clearer,so super-resolution reconstruction comes into being.At the same time,in recent years,deep learning has become more and more widely used in various fields due to its powerful learning ability,Image super-resolution reconstruction based on deep learning has also become an important research direction in the field of computer vision.This paper focuses on the research of image super-resolution reconstruction method based on deep learning,the main work is as follows:1.The relevant theoretical knowledge of deep learning in the field of image super-resolution is systematically explained,including convolutional neural networks,residual networks,dense networks,learning-based upsampling methods,loss functions,etc.At the same time,the image super-resolution reconstruction methods are classified,the classic algorithms in different methods are explained,and the advantages and disadvantages of traditional methods and deep learning methods are compared.Finally,we introduce the relevant indicators of image quality evaluation.2.An image super-resolution reconstruction method based on multi-scale dense residual network is proposed.In order to solve the problem of insufficient feature detail information extracted by the single-scale convolution of the mainstream networks at present,a multi-scale convolution network is designed in this paper.By integrating the features of different scale convolutions,the information of the reconstructed image is richer.In addition,in order tostrengthen the mobility and reuse of features,and to avoid the disappearance of gradients during training of deep networks,this paper integrates densely connected networks and residual networks into multi-scale convolutional networks to make the network's performance stronger.3.An image super-resolution reconstruction method based on multi-scale pyramid network is proposed.Since the current mainstream network using only one upsampling during the image reconstruction process,this will make it more difficult to train a high scale factor(such as 8 times).In order to solve this problem,a multi-scale pyramid network is designed in this paper,and the rich features extracted by the multi-scale residual module are gradually up-sampled in a pyramid manner,so that the reconstruction effect of the network for high-scale factor images is further improved...