Research On Unsupervised Image Generation Based On Generative Adversarial Networks

With the advent of big data era,the difficulty and complexity of data analysis have dramatically increased.There are some limitations in the way of obtaining data information,which completely depends on the traditional statistical model.Especially for the analysis of unstructured data,the limitations are particularly evident.Therefore,how to analyze valuable information from massive data is a major challenge in the current statistics field.It is an effective way to address complex statistical problems by breaking through the limitation of field and developing the statistical model of multi-discipline integration.Generative model is one of the most important statistical models,which mainly leverages the real data distribution to establish the model and solves the problem of sample generation.In this study,we focus on a deep probabilistic generative model: generative adversarial networks,and its application in image generation.Due to the excellent effect in generating unstructured data such as image,speech,text and video,generative adversarial network has quickly become one of the mainstream deep generative models,which has been widely used in many discipline fields,such as computer science,medicine,biology.However,generating high-quality samples is still one of the challenging issues in this field,especially in unsupervised scenario(without any label information).To this end,in this dissertation,we systematically study the sample generation problem of generative adversarial networks in unsupervised scenarios from image generation perspective.The contributions of this dissertation are summarized into three aspects as follows.Firstly,an effective algorithm of generation adversarial networks is proposed based on information entropy.Specifically,we leverage conditional entropy to construct a statistical distance in a strict sense,that is,we prove that the distance satisfies three conditions in metric space: identity of indiscernibles,symmetry and triangle inequality.Then,this distance is directly penalized on the target function of the generator,the purpose of that is to force the generated distribution to sufficiently approximate the real distribution by this distance.The experimental results show that the algorithm can significantly improve the quality of sample generation.Secondly,this dissertation also studies the training instability for generative adversarial networks.In this study,we first introduce a generalized generator loss function,and then analyze the causes of network training instability,and conclude that the key causes of network training instability lie in: when the generator converges to the minimum,the discriminator cannot converge to the optimal discriminator;While when the discriminator approaches the optimal discriminator,and the gradient update of the generator is unstable.According to this fact,we find that controlling the Lipschitz constant of discriminator is a breach to deal with the problem of network instability.Along this line,we propose a zero-centered gradient penalty algorithm to address this problem.The experimental results show that the algorithm can not only make the network training stable,but also make the network obtain good convergence.Finally,in order to address the problem of mode collapse of generative adversarial networks,an integrated algorithm for training generative adversarial networks is proposed by coupling spectral normalization with the zero-centered gradient penalty technique.This algorithm integrates the advantages of two technologies,which has a good effect on addressing the mode collapse problem of generative adversarial networks,and significantly improves the diversity and fidelity of the generated samples.The algorithm has a significant advantage in unsupervised scenario compared with the generated sample quality of several representative algorithms.In addition,we apply this integrated algorithm to the classification problem of imbalance data.Specifically,the generated samples by the algorithm are added into the minority-class samples,reduce the imbalance of data,and improve the classification accuracy.The experimental results show that the algorithm can effectively deal with the classification problem of imbalance data.