Research On Key Technologies Of Generative Adversarial Network Generation Model

The generative adversarial network(GAN)is an important basic model in artificial intelligence research field.It mainly consists of a generator and a discriminator to construct a two-player zero-sum game adversarial learning model,and completes adversarial learning through sequential training of the discriminator and generator.GAN can simulate the implicit distribution of training data,in the visual performance,the generator can generate some random samples similar to the training data.In recent years,the extended application tasks of GAN are very wide,such as image inpainting,image translation,super-resolution,style transfer and objective detection etc.The research of GAN related key technologies is an important basis for its application.In this dissertation,three kinds of key technologies related to the GAN-generated image quality are studied: the training method,loss function and adversarial learning stability techniques.To explore the application value of the previous research works in this dissertation,SFDC GAN,ANL-CG GAN and MPF-GAN related technologies are applied to high magnification face super-resolution(FSR)to improve the human visual quality of the FSR image.The main works of this dissertation are as follows:1.In DCGAN,to improve the generated image quality of small batch random sampling training method,a DCGAN training method based on subsample set construction is proposed.First,a feature space based on the first order color moment and clarity is constructed,and the training samples are projected into the feature space.Second,the probability density function of the training samples in the feature space is calculated to replace that of the training samples.Finally,several subsample sets are constructed by the improving probability sampling method,which makes the subsample sets approximately independent and identically distributed and closer to the training sample distribution.The experimental results show that when the Batch Size is 2048,the novel training method can improve the generated image diversity and quality.2.In the adversarial learning optimization function,the common problem in the optimization process is the vanishing gradient problem when optimizing JS divergence.Aiming at this problem,this dissertation proposes a GAN model based on decoding constraint of training sample feature(SFDC GAN).First,the autoencoder with similar structure of U-Net is used to learn the middle layer features of samples with the same dimension as the random vectors used in the generator.Second,the decoding constraint is designed,which is used to train decoder before each adversarial training.The decoder has the same structure and weight sharing with the generator,so that the decoding can constrain generator parameter update.The experimental results show that the novel GAN can improve the generated image quality,which is better than classical GAN and close to SAGAN.3.The outlier noisy data of training samples,the balance of game ability between generator and discriminator are two factors that affect the adversarial learning stability.To improve the adversarial learning stability,the antinoise learning and coalitional game GAN,ANL-CG GAN is proposed.In ANL-CG GAN,it achieves the above goals through the following two strategies.(1)In the real sample loss function of the discriminator,an effective antinoise learning method is designed,which can improve the gradient variance and network convergence uncertainty.(2)In the two-player zero-sum game,a generator coalitional game module is designed to enhance its game ability via coalitional game strategy.The coalitional game module can improve the balance between the generator and discriminator.The experimental results show that the ANL-CG GAN can improve adversarial learning stability,reduce the number of training epochs and improve the generated image quality.It is better than classical GAN and close to SAGAN and SNGAN.4.The GAN optimization problem of two-player zero-sum game is usually a single objective unconstrained optimization problem of nonconvex optimization,which is easy to fall into local optimal solution and difficult to control the adversarial learning stability.For the above problem,to improve the adversarial learning stability,a multi-penalty functions GAN via multi-task learning,MPF-GAN is proposed.First,the single objective unconstrained optimization problem is transformed into a single objective constrained optimization problem with the same solution.Its constraints are two equality constraints that can be implemented by adversarial learning.Second,an augmented objective function with penalty function is constructed by using the penalty function method.Finally,the multi-task generator and multi-task discriminator GAN network are constructed by CNN and Res Net,and these penalty functions and single objective function are trained in turn to train the adversarial learning.The experimental results show that the MPF-GAN can improve adversarial learning stability and the generated image quality.It is better than classical GAN and close to SAGAN and SNGAN.5.In parameter updating of adversarial learning generator and discriminator,because the parameter updating is complex and changeable,directly optimizing a single divergence is easy to lead to local optimal solution,which can affect the adversarial learning stability.For the above problem,to improve the stability,a same solution constraint GAN,SSC-GAN is proposed.First,the probability density function of training samples is reasonably assumed,and a Cauchy initial value problem of first-order ordinary differential equation(ODE)is constructed based on this probability density function.The ODE is also the same solution problem of GAN optimization problem,and the constraints on the existence and uniqueness of the ODE solution are given.Second,adversarial learning is regarded as solving the ODE,and these constraints are incorporated to ensure the existence and uniqueness of GAN solution as much as possible,so as to improve the adversarial learning stability.The experimental results show that the SSC-GAN can improve adversarial learning stability and the generated image quality.It is better than classical GAN.It is also better than SAGAN and SNGAN.6.Direct use of end-to-end learning for high magnification(such as 8(?))face super-resolution(FSR)tasks,the FSR image is fuzzy and the human visual quality is poor.To improve the human visual quality(photo-realistic)of FSR image,the adversarial learning method is introduced.In End-to-End learning,to improve the FSR image quality,using decoding adversarial learning of SFDC GAN(Chapter 4)and multi-task adversarial learning MPF-GAN(Chapter 6)technologies,a deep multi-task Laplace pyramid network is designed.The main-task is End-to-End learning,the sub-task is the penalty function of adversarial learning.In the End-to-End network parameter tuning,to improve the FSR image quality after parameter fine-tuning,antinoise adversarial learning is integrated via ANL-CG GAN(Chapter 5)antinoise learning technology.The experimental results show that the proposed FSR algorithm can make the super-resolution image more photo-realistic and more in line with human visual habits.