Research On Domain Generalization By Feature Style Augmentation

Although deep learning systems have achieved remarkable performance in many fields,the problem of serious performance slippage caused by domain shift still needs to be solved.Domain generalization methods can improve the model’s generalization performance to arbitrary unseen domains by simultaneously learning the invariant paradigm from multiple domains.Recently,the MixStyle method proposed by Kaiyang Zhou et al.generates more diverse stylized features through linear interpolating convolutional feature statistics,significantly improving the model’s generalization performance in unknown fields.Therefore,this thesis further studies the domain generalization method from the perspective of feature style augmentation.In addition,this thesis studies how to use semi-supervised and active learning methods to solve the problem of domain generalization in data sparsity.The main work of this thesis includes:1.This thesis proposes the domain generalization method of SDFA(Style Decoupling Feature Augmentation).In feature style augmentation,SDFA processes the mean and standard deviation of convolutional feature statistics separately by performing entirely random linear interpolatiaon,thus achieving a more diverse feature style augmentation.Experimental results show that the average accuracy of SDFA is improved by 1.2%compared with MixStyle on the PACS dataset.On the OfficeHome dataset,the average accuracy of SDFA is 0.2%higher than the MixStyle method.2.This thesis proposes a domain generalization method named FSADG(Feature Stylization Adversarial Domain Generalization).FSADG inserts a style augmentation module based on VAE(Variational AutoEncoder)between the model’s convolutional layers or residual blocks.The augmentation module learns the latent distribution of feature styles and obtains new stylized features by adding noise to latent vectors and adversarial learning.Experimental results show that the average accuracy of FSADG is 3.1%higher than MixStyle on the PACS dataset.On the OfficeHome dataset,the average accuracy of FSADG is improved by 0.6%compared with MixStyle.3.For the domain generalization problem in data sparsity,this thesis proposes a semi-supervised learning-based domain generalization framework named SSDG(Semi-Supervised Domain Generalization).SSDG provides a solution to the domain generalization problem of small samples by combining Fixmatch with FSADG.The experimental results show that on the PACS dataset,SSDG achieves an average accuracy of 79.1%using only 105 tag samples,much higher than the average accuracy of 54.5%of MixStyle and slightly lower than the average accuracy of 79.5%in the case of total sample supervised learning.4.Based on Work 3,this thesis studies how to use active learning to expand the original labeled training set efficiently.This thesis proposes an active learning sampling strategy named DAS(Domain-agnostic Sampling)for domain generalization.DAS assumes that the samples that the current domain discriminator cannot clearly distinguish are the most valuable for labeling.The experimental results show that the sampling efficiency of DAS is higher than that of the classical ME(Max Entropy)sampling algorithm in the small sample domain generalization scenario.