Research On Image Super-Resolution Based On Generative Adversarial Networks

Image super-resolution reconstruction methods have been widely applied in satellite remote sensing,medical diagnosis,public security,and so on.With the rapid development of computer science and technology,more and more single image super-resolution reconstruction methods have been proposed.Existing super-resolution reconstruction methods mainly use the deep network to learn the mapping relationship between the low-resolution image and high-resolution image,and then reconstruct the corresponding high-resolution image.In order to improve the reconstruction performances of the network,many researchers extend the depth or width of the network.But it makes the network redundant and increases the computation cost.To avoid designing the network blindly and save the number of network parameters,this thesis studies the super-resolution reconstruction methods under the guidance of expert priors from the perspective of model-driven and data-driven.It proposes an enhanced image super-resolution method using hierarchical generative adversarial network,a recursive dense residual generative adversarial network based super-resolution,and an edge-driven generative adversarial network method.Our main works are as follows:(1)Existing image super-resolution reconstruction methods based on generative adversarial network have simple network structure,and often ignore the importance of prior information in image restoration.Therefore,this thesis introduces the prior knowledge into the generative adversarial network,and proposes an enhanced image super-resolution method using hierarchical generative adversarial network for building more realistic high-resolution images.Firstly,the generator network of our method adds edge extraction branch and edge enhancement branch beside the image reconstruction branch.The edge extraction branch extracts highfrequency information in the low-resolution image ignored by deep network,while the edge enhancement branch extracts sharp edge information and compensates the edge extraction branch.In the meanwhile,the edge extraction branch and edge enhancement branch are fully trained by adding the edge loss in the loss function.Then the features extracted from above two edge branches are fused to the image reconstruction branch.Finally,with the gambling of the generator network and the discriminator network,a satisfying reconstructed high-resolution image is obtained.The experiments illustrate this method of better reconstruction performances.(2)Although deepening the network depth can improve the network reconstruction performance to a certain extent,it will also bring a large number of network parameters,which makes network training difficult.In order to reduce the number of network parameters and ensure the reconstruction performance as much as possible,this thesis combines the idea of recursion with expert prior,and proposes a super-resolution reconstruction method based on recursive dense residual generative adversarial network.Firstly,the image reconstruction branch of the generator network is composed of dense residual blocks of recursion,and the feature information obtained after each recursive operation is fused.Secondly,the edge extraction branch of the generator network is constructed by using shallow network.In the end,the image reconstruction branch and edge extraction branch are fused to obtain a high-resolution image with clear edges.Experimental results show that the recursive operation in this method can achieve the same or better reconstruction effect without increasing the number of network parameters.(3)In order to improve the reconstruction performances,many existing methods based on the generative adversarial network focus on deepening the generator network or discriminator network.But it will make the reconstructed image lose many high-frequency details,and also increase the memory and the computing cost.For emphasizing the edge details and simplifying the network structure,this thesis proposes an effective method,named edge-driven super-resolution generative adversarial network,by fusing the idea of data-driven and model-driven.In the generator network of our proposed method,an edge extraction branch is considered besides the reconstruction branch.But different from the traditional edge extraction branch that uses a deep subnetwork blindly,we design it under the guidance of an iterative algorithm deduced from the edge prior.By means of the added edge loss and the sharpness loss in the traditional loss function,the generator will be trained well to generate sharper high-resolution images.Experimental results illustrate that our proposed method can not only improve the image reconstruction performances effectively by using the edge prior,but also avoid deepening the network blindly.