The Research And Application On Generative Adversarial Networks Oriented Towards Computer Vision

With the rapid development of artificial intelligence and deep learning,research on computer vision-oriented generating models has been widely used.Generative Adversarial Networks(GANs)among them have the most far-reaching impact.Compared with other traditional generating models,the adversarial training between the generator and the discriminator in GANs can significantly improve the quality of generated samples while avoiding the derivation process of approximation or variation of traditional generating models.On the other hand,the generator does not directly participate in the learning of the data distribution,which makes generative adversarial networks insensitive to the prior knowledge related to the data distribution,so that new samples similar to but not identical to the data distribution can be generated better.Although generative adversarial networks have many advantages and have been successful in different tasks,there are still two inherent drawbacks: training dynamic instability and mode collapse.This dissertation combs the related works and the development status of GANs,analyses the causes of both two inherent drawbacks according to the manifold theory and the optimal transport theory,and proposes two solutions from different perspectives.On this basis,the application of generative adversarial networks in semisupervised learning and reinforcement learning is further expanded,which provides new ideas for the application of generative adversarial networks in the medical field.In this process,the effectiveness of the proposed method and applications are fully verified by theoretical analysis and large-scale experiments.A detailed overview of this dissertation is as follows:(1)Directing against the two inherent drawbacks of generative adversarial networks,this dissertation proposes a framework named CES-GAN for GANs based on the ensemble of multiple discriminators with different cost functions.It can capture the observable data modes in the data space as much as possible and stabilize the training dynamic of GANs by the ensemble training of multiple discriminators with different cost functions simultaneously,benefiting from the characteristics of different cost functions that have different preferences on data modes.Meanwhile,in response to the Cannikin Law problem in the ensemble of discriminators with different performances,this dissertation also proposes a gradient selecting mechanism,which only uses gradients that can effectively decrease the distance between the data distribution and the generated distribution to update weights of the generator,so as to avoid the negative impacts of the discriminator with lower performance.Finally,this dissertation provides theoretical analyses to prove the reliability of the proposed method and verifies the effectiveness of the method on multiple datasets with different scales.(2)Aiming at the problem of unstable training dynamics of GANs,this dissertation proposes an express construction method for them based on latent representation.The generation process of GANs is divided into two independent generation processes to generate empirical latent representation and final results respectively,and thus stabilize the training dynamics of GANs and capture more data modes.In the second stage,the empirical latent representation is used to replace the white noise as the input of the generator.Considering that the latent representation is a low dimensional mapping of the data distribution in the latent space of the neural network,with prior knowledge of the data distribution,means that their mapping functions are more continuous and smooth.Therefore,this method can effectively stabilize the training dynamics of GANs.Finally,this dissertation proves that the proposed method can effectively stabilize the training dynamics of GANs and reduce the difficulty of its training through theoretical analyses.Large-scale experiments on several datasets demonstrate the effectiveness of the method.(3)In the sight of the over-fitting problem caused by the imbalance between general and specific features of the deep learning classification network in supervised learning,this dissertation proposes a semi-supervised training method based on feature equilibrium.Convolutional layers at the top of the classification network are regarded as the generator,and its output is taken as fake samples.The projections of real samples obtained through matrix transformation are viewed as true samples.An unsupervised discriminator is introduced to distinguish the two samples based on an adversarial loss,thereby constraining the model to learn more general features rather than specific ones.Theoretical analyses and large-scale experiments show that the method is reliable and effective.(4)Pointing at the problem that the radiological features of COVID-19 are too similar to those of other pneumonia,this dissertation proposes a semi-supervised learning detection model based on GANs,to identify the subtle differences between CT images of different types of pneumonia.Meanwhile,given the inductive bias problem caused by the imbalance between radiological features and clinical indicators,this dissertation also proposes a reinforcement learning prediction framework based on GANs to predict the progress of patients with COVID-19 using both radiological features and clinical indicators.This dissertation also constructed a chest scanning image dataset of COVID-19,containing more than 100 thousand CT images,and carried out a large-scale verification on this basis.Besides,the tests on external verification sets prove the effectiveness and generalization ability of the proposed method.The methods proposed in this dissertation can significantly stabilize the training dynamics of GANs and capture mode data modes.At the same time,the application results on semi-supervised learning and reinforcement learning show that the GANs have high potential application value.