Unsupervised Disentanglement For Generating Adversarial Networks And Its Optimization

In recent years,the research and application of generative adversarial networks have made great progress.However,during the training process,the network model often fails to map the input noise to the real data distribution,causing the model to collapse and the generated attributes to be entangled,making it difficult to achieve satisfactory quality and controllability of the generated content.With the in-depth research in the industry on optimizations such as disentanglement and resistance mode collapse of generative adversarial networks,unsupervised disentanglement schemes and optimization from the theory itself have gradually become hot spots.This dissertation revolves around the optimization of generative adversarial networks for unsupervised disentanglement and resistance to mode collapse.The specific content is divided into two aspects.On the one hand,by improving the unsupervised disentanglement method of the generative adversarial network,the attributes of the generator can be better separated to achieve the purpose of improving the quality of generation and controlling the attributes;Combined with game theory,the training algorithm of the generative adversarial network is improved to achieve the purpose of alleviating mode collapse and improving the performance of network generation.Work on improving unsupervised disentanglement of generative adversarial networks focuses on enhancing model attribute separation.Various unsupervised disentanglement works are excellent in shallow attribute separation,but the performance of disentanglement in the case of deep content and too many attributes is average.To solve this problem,firstly,by analyzing the principle of disentanglement of the generative model by Hessian matrix and Jacobian matrix at different levels,and then further deduce the inner relationship between the two,and use different difference approximation to make a comparison between the two matrix methods.The optimization process is computed to derive an unsupervised consistent disentanglement loss,which is applied to disentanglement between continuously varying factors and between multiple varying factors.The experimental results show that the method proposed in this disertation has certain advantages compared with multiple current excellent disentanglement algorithms under multiple data sets,and has better performance under the test of disentanglement generation with the separation of various attributes.In terms of the optimization of generative adversarial networks,the main work is to alleviate mode collapse and reduce generated artifacts and semantic ambiguity.This dissertation improves the hawk dove game model,modeling and analyzing the generative adversarial network by analogy,and modifies the model architecture of the generative adversarial network while finding an improved theoretical solution.In this disertation,the game relationship between the generator and the discriminator is extended to the asymmetric game,and a new generation training paradigm is added without changing the original model as much as possible.Experiments have proved that after adding multiple mainstream generative adversarial network models into the optimization method of this disertation,most of the models generated perform better than the original model,and the quality and diversity of generated images are mostly better than the methods before improvement.Figure [19] Table [7] Reference [82]...