Research On Quality Improvement Technology For Generative Adversarial Network

In recent years,generative adversarial networks in deep learning have gradually become the mainstream of generative model algorithms.Since the generative adversarial network was proposed in 2014,there are already thousands of improved models that shows the important position of generative adversarial networks in the field of generative models.Generative adversarial networks have been widely used in different fields.For instance,generating various types of data such as images,audio,and video.The various derived models of generative adversarial networks have also played an important role in many important areas,such as natural language processing and pattern conversion.Thus,the quality improvement technology of generative adversarial networks has attracted more and more attention from researchers at home and abroad.In the existing research on generative adversarial networks,most of the studies merely focus on the loss function and the structure of the network model.Whereas few studies focus on the improvement of the internal structure and training methods of the network,this paper aims to improve the quality technology of generative adversarial network by optimizing the internal structure and training methods.Firstly,this paper will study the comprehensive quality improvement technology of generative adversarial network based on precision optimization for the comprehensive quality problem of generative adversarial network in existing research.Next,this paper will study the efficiency improvement technology of generative adversarial network in the iterative scenario of generative adversarial network in existing research.Finally,this paper will study the model construction technology of generative adversarial network that supports external feedback and dynamic adjustment for generative adversarial network external feedback and training stability problems in existing research.The relevant research work have achieved the following innovative results:(1)In this paper,aiming at the comprehensive quality problem of generative adversarial network in existing research,we propose the comprehensive quality improvement method of generative adversarial network based on precision optimization.Although there are many implementation models for multi-objective generative adversarial networks in the existing literature,for example,ACGAN separates the output of the fidelity and category results of the discriminator,and VACGAN separates the realism and category networks.Neither of them have carefully taken the network structure into consideration.Also,the comprehensive accuracy of the multi-objective generative adversarial network has not been studied in depth.Based on the original model,this paper optimizes the internal structure of multi-objective discriminant network to improve the comprehensive quality of the network.More specifically,this paper will consider the correlation between multiple targets,focus on the network internal structure of the discriminator in the multi-objective generation adversarial network,and integrates multiple optimization targets to optimize the internal structure of the network,such as multi-objective optimization of classification and realism.We can neither let multiple targets influence each other too much,nor can we make them irrelevant.Firstly,this paper observes improvement space of the multi-objective comprehensive quality of the generated adversarial network.Then,in view of the problem of the interaction between multi-objective discriminators,this paper proposes a partial splitting method to generate multi-objective discriminators in an adversarial network,and gives the parameters of the method,so that the training of multiple targets can promote each other without affecting each other.Next,aiming at the problem with not being able to train properly for partial split method of multi-objective discriminator,this paper proposes a multi-objective discriminator differentiation training method,which enables the model to better adapt to various improved versions of generative adversarial networks.Finally,this method conducts experiments on large-scale signaling data,and selects multiple models as comparative models.Experimental results show that the proposed method effectively improves the comprehensive quality of generative adversarial networks.(2)In this paper,we propose the efficiency improvement method of generative adversarial network in the iterative scenario of generative adversarial network in existing research.Most of the recent studies focus on the quality of data generated by generative adversarial networks,but there is no further research on the efficiency of generative adversarial networks.Based on the original model,the new model proposed in this study optimizes the internal structure of the generative network or discriminant network to improve the efficiency of network training.Firstly,this paper observes the efficiency requirements of generating adversarial networks in iterative scenarios,and analyzes the problems that require rapid convergence of generative adversarial networks.Then,aiming at the problem of waste of training resources of the same feature by multiple discriminators,this paper proposes a data feature splitting method based on width calculation.This enables the data to be separated into small implicit features.Next,aiming at the training problem of data feature splitting method based on width calculation,this paper proposes a multi-discriminator training method with controllable discriminant strength.In this way,the generative adversarial network can produce better results faster.Finally,this method conducts experiments on data such as CIFAR-10 and anime avatars,and selects multiple models as comparative models.The experimental result shows that this method could effectively improves the training efficiency of generative adversarial networks.(3)In this paper,we propose the model construction method of generative adversarial network that supports external feedback and dynamic adjustment in view of the stability of generative adversarial network training in existing research.We found that most of the existing research focus on optimizing the stability of generative adversarial networks in perspective of the loss function.However,there is no research on the direction of adding feedback interventions to the network structure.The new model proposed in this study,based on the original model,optimizes the internal structure of the generative adversarial network and adds human intervention to improve the stability of the generative adversarial network in the training process.Firstly,this paper observes that the adversarial indicators in the process of network training are so unstable that model crash or quality collapse often occurs.Thus,the stability problem of generative adversarial network training is analyzed.Because the stack generation adversarial network model produces rich external feedback,in this study,we choose the stack-generated adversarial network as the base model.Next,this paper proposes a training method that supports external feedback and dynamic adjustment for generating adversarial networks using external feedback tuning.At last,this method conducts experiments on CIFAR-10 data and selects multiple models as comparative models.The experimental result shows that this method could effectively stabilizes the training process of generating adversarial networks and improves the training quality.