Research On Image Super-resolution Algorithm Based On Deep Learning

In recent years,the research of image super-resolution has made great progress.In general,compared with low-resolution image,high-resolution image can show fuller tex-ture and more obvious edge,so people can get more information which is helpful for image understanding.Therefore,the high-resolution image is more conducive to the subsequent decomposition,processing and application of the image,and also can improve human visual enjoyment.Image super-resolution is mainly to restore image details through low-resolution images and reconstruct corresponding high-resolution images.Generally speaking,most satisfactory results are achieved by very deep models,which follow a simple principle that the deeper the better.However,blindly increasing the number of network layers will inevitably lead to problems such as large amount of network parameters,long calculation time,over fitting,gradient disappearance or gradi-ent explosion.At the same time,in the fully-connected convolutional neural network,the information of each layer can only be transmitted to the upper layer,and the information processing at different time is independent,so the influence of time on the neural network is not considered.In addition,the traditional feed-forward structure is difficult to make full use of the interdependence between low-resolution and high-resolution images.Based on these observations,for a single natural image,a recurrent neural network with two states is proposed by using recursive learning,which makes this model lightweight and efficient.On the basis of dual-path block,the network performance can be improved by increasing the number of dual-path block without adding additional parameters.In ad-dition,compared with most single-state convolutional neural network methods,the pro-posed network has two states(low-resolution and high-resolution)which are transformed to each other.With back-projection,the method makes full use of the dependency between the two states and calculates the difference between the two states to reconstruct better re-sults.A large number of experiments show that this method can reconstruct images of better quality.In addition to the problems mentioned above,ground targets in remote sensing im-ages usually have large scale,which means that the joint distribution of image patterns be-tween the target itself(such as aircraft)and its surrounding environment(such as airport)is coupled,which is different from natural images.For remote sensing satellite images,this thesis also proposes a new super-resolution method,which can closely connect the traditional recurrent neural network and channel attention mechanism.This method pays enough attention to the high-frequency channel information through recursive channel at-tention block,and reduces the low-frequency information adaptively.This method uses the correlation between the low-resolution and high-resolution states to learn more and more high-level abstract feature expression.This method also uses the hierarchical feature fusion of low-resolution and high-resolution to reconstruct the high-resolution image with high-frequency details.Recursive channel attention block improve network performance without adding additional parameters.