Research On Application Of Image Content Generation And Unsupervised Domain Adaptation Technology

This dissertation focuses on the application of image content generation and un-supervised domain adaptation.The main research contents include image content syn-thesis based on deep generative models and unsupervised domain adaptation method.The first part of this dissertation is image content synthesis based on deep generative models includes two parts:natural texture synthesis based on variational auto-encoder and remote sensing image restoration based on neural autoregressive model.(1)A tex-ture synthesis method based on image patch is proposed.The variational auto-encoder is used to learn the feature representation of natural texture,and the new texture is generated by interpolation sampling in the latent variable space.(2)Aiming at the problem of image degradation caused by cloud occlusion in optical remote sensing image,this dissertation proposes a method to remove cloud occlusion based on neural autoregressive model.This method is called Cloud-Aware Generative Network,which mainly includes two sub networks:cloud occlusion detection and cloud occlusion re-construction.The Cloud-Aware Generative Network can remove cloud cover by using a single frame of optical remote sensing image without using multi temporal information.The detection sub network of cloud occlusion region introduces attention mechanism to assist potential cloud occlusion region segmentation,and partial convolution layers are embedded in the reconstruction sub network of cloud occlusion to shield the noise caused by cloud occlusion.Experiments are carried out on synthetic cloud occlusion remote sensing image and real cloud occlusion remote sensing image.The experimental results show that the peak signal-to-noise ratio and structural similarity index of cloud occlusion image processed by the Cloud-Aware Generative Network can be significantly improved.The second part of this dissertation focuses on the unsupervised domain adaptation method.The goal of unsupervised domain adaptation is to improve the generalization performance of the model trained with annotated source domain data in unlabeled target domain data.In the scenarios of unsupervised domain adaptation,the learn-ing paradigms suffer from the domain shift.Many methods reduce the distribution difference between the source domain and the target domain through the probability distribution alignment of the source domain data and the target domain data.How-ever,many previous unsupervised domain adaptation methods based on distribution alignment ignore the use of latent global structure information of categories,which can be used as an effective framework to guide cross domain knowledge transfer.In this dissertation,an unsupervised domain adaptation regularization method is proposed,which is called Global Clustering Central Structure Regularization.This method guides the model to migrate from the source domain to the target domain by constraining the consistency of the class center structure of the source domain data and the target domain data.Specifically,this paper constructs a topological graph structure about clustering centroid,and regularizes the cross domain feature extractor by enhancing the consis-tency of the graph in the two domains.The proposed method matches the first-order and second-order proximity of the clustering center graph,which can complement several new unsupervised domain adaptive methods.The Global Clustering Central Structure Regularization method can be combined with the existing domain adaptation methods only by using a simple additional method.In order to prove the effectiveness of Global Clustering Central Structure Regularization method,we systematically evaluated the method on three unsupervised domain adaptation benchmark datasets,and compared it with a series of state-of-the-art unsupervised domain adaptation methods.The results show that the global cluster central structure graph inherits meaningful cross domain invariant knowledge,which is conducive to cross domain transfer learning.In this disser-tation,the general theory and technology of image generation are carried out for texture synthesis and remote sensing image restoration,and a new regularization method is proposed to improve the quality of distribution adaptation in unsupervised adaptation tasks,which has important theoretical significance and engineering practical value.