Research On Generative Adversarial Network And Its Applications

In recent years,the rapid developments and wide applications of deep learning have led us to appreciate the great potential of artificial intelligence to bring about changes in people's life.As a generative method,the generative adversarial network(GAN)is trained against two deep neural networks and optimized with stochastic gradient descent,which avoids not only the calculation of the partition function induced by the repeated applications of Markov chain learning mechanism,but also the need to approximate inferences.Thus it greatly promotes the applications and has far-reaching significance for the development of the generative model.GAN has been widely studied and paid close attention by the academic and industrial circles of artificial intelligence immediately after its proposal.With the development of deep learning technology,GAN has been continuously promoted in theory and application.At present,GAN has been applied successfully in the fields of computer vision,speech and natural language processing.Though GAN overcomes some problems of traditional generative methods,it also brings about new issues.In training process,it is difficult to guarantee that the two competing networks can well balance and synchronize,resulting in the instability of training GAN.Besides,the mode collapse problem and the poor interpretability that stems from neural networks are also general to GAN.Settling of these problems plays a key role in promoting the applications of GAN.The purpose of this dissertation is to explore GAN under different learning paradigms and apply these theoretical researches to video monitoring of flow velocity.1.Aiming at the shortage of supervised GAN,we propose an improved supervised conditional GAN.We use the unpaired images in target domain for the conditional variable of the discriminator,thus to effectively control the generating process by full use of conditional variables as well as to provide more prior information than an exact target image dose.In addition,the blurring in GAN induced by pixel-wise reconstruction can be effectively overcome by the binary feature matching,which additionally takes the depth hierarchical features into account besides the pixel-wise features.On this basis,a supervised general architecture suitable for image processing,such as image inpainting,semantic segmentation,image colorization and de-noising,is proposed.2.In order to make GAN applicable to the ubiquitous semi-supervised application scenarios,that is,a few labeled samples while a large number of unlabeled ones,a semi-supervised learning mechanism is designed,which sets up special channels for supervised and unsupervised samples.On the supervised channel,the binary features matching proposed in the third chapter is used for the supervised training,so as to capture the visual features and the intrinsic structural characteristics of the image in the target domain simultaneously.Acting as a prior,the real images are introduced to better guide the generation process.3.Precise clustering is an essential part of successful semi-supervised GAN.Therefore,this paper further proposes a novel clustering method suitable for semi-supervised GAN.It adaptively learns the feature weights and fits the feature selection and clustering into a unified framework,rather than the two-phase strategy of typical approaches.With a new rank constraint imposed on the Laplacian matrix,the connected components in the resulted similarity matrix are exactly equal to the cluster number.4.In order to improve the stability of unsupervised training process of GAN,a closed loop from the source domain to the target domain and the cross domain learning from the target domain to the source domain is constructed based on the dual learning,so that GAN can carry out effective unsupervised learning even in the absence of manual intervention and expertise.5.The applications of GAN under the different learning paradigms proposed in the previous sections are applied to the video monitoring of flow velocity,which are not only the validation of the proposed model,but also the extension of the application of GAN.We use the supervised NPCGAN to denoise the flow images with interference such as dense rain streak,snowflake or thick fog,and the unsupervised DADualGAN to automatically lighting the flow images collected under the condition of inadquate light.Based on the relationship between the wave characteristics of the water surface and the velocity of the flow,the semi-supervised LRAFLGAN is used to identify the velocity range of the real-time water surface image,so as to realize the flow monitoring with low cost and high reliability without delay.To sum up,the main contribution of this dissertation is the research on the theory and application of GAN,which is one of the latest and practical deep learning methods.We explore the GAN under various learning paradigms,namely the supervised,semi-supervised and unsupervised learning.Technologies of non-parallel conditional variable,binary features matching and so on are proposed.By introducing the domain adaptation and dual learning into the training mechanism of GAN,both the instability of training process and the generation quality of GAN are effectively improved.In addition,it also tries to extend the GAN to the application of real-time flow monitoring in the intelligent scheduling system.