Generative Adversarial Networks And Its Application Research

Generative adversarial network is a kind of generative model in deep learning.It can learn data distribution and obtain new samples by sampling.In the field of image processing,generative adversarial networks have applications in many scenarios,such as data expansion,autonomous driving,expression conversion,and artistic creation.But it still has problems such as mode collapse,training collapse,and lack of robustness of the model.Therefore,the research and generative adversarial networks in image processing issues can help to improve its commercial application value and promote the development of artificial intelligence.This paper mainly studies the performance improvement of generative adversarial networks,the application of generative adversarial networks in multimodal style transfer,and the existence of adversarial samples in generative adversarial networks,mainly through theoretical analysis and experimental verification.The system of this paper includes the performance improvement of generative confrontation networks,application research and application robustness research.The main innovations include the following three aspects:1.In response to the problem of mode collapse of the generative adversarial network,we propose a multi-mode generative adversarial network based on the latent Dirichlet distribution,which can discover the inherent hidden structure of the data and has a significant improvement in fitting the data distribution.The experiment has verified the structural information learned by the model,and has greatly improved the indicators of evaluating quality and diversity.2.For the application problem of generative adversarial networks,we conduct research in the field of multimodal style transfer,and there is also a mode collapse problem in this field,which will lead to insufficient diversity of pictures after migration.We propose adaptive EM Attention mechanism to solve this problem,further development in learning style characteristics.In addition,the feature layer is visualized through experiments,which proves the effectiveness in style learning,and also has a significant improvement in the evaluation of diversity indicators.3.For the robustness problem of generative adversarial networks,we conducted a study of adversarial samples in the field of multimodal style transfer.The adversarial samples have important significance in the field of security and model robustness,but in multimodal style transfer There is still a lack of research.We first give a definition of this problem and propose a corresponding solution.In addition,we also proposed an innovative loss function,which has a good effect in quantitative analysis.