The Research On Multi-source Domain Generalization For Image Classification

Modern deep neural networks suffer from performance degradation when evaluated on testing data under different distributions of training data.Domain generalization(DG)aims to tackle this problem by learning transferable knowledge from multiple source domains so as to generalize to unseen target domains.However,trained models perform well in one dataset,only to perform poorly in another dataset,owing to the varying bias within the DG datasets.Therefore,this paper conducts a study on multi-source DG for image classification to address the different domain shift in datasets.The main works are listed below:(1)Aiming at the problem of correlation shift in DG datasets,this paper proposes a multi-source DG method based on dual variance invariance constraints.Correlation shift is caused by spurious correlations of invariant features and labels in the training domain.Therefore,the inter-domain variance risk loss is introduced to estimate the risk error in each domain,so that the variance of the average risk loss between training domains is minimized,thereby reducing the discrepancy between domains.Subsequently,inspired by the information bottleneck theory,feature entropy loss is applied to compress the learned representations so that the influence of false features on labels is eliminated as much as possible.Finally,experiments are conducted on the Colored MNIST dataset with correlation shift,achieving an accuracy rate of 69.9%.(2)Aiming at the problem of diversity shift in DG datasets,this paper proposes a multi-source DG framework based on data augmentation and intra-domain contrastive learning.Style differences between domains can lead to diversity shifts,thus,this model aims to reduce this effect.First,a novel Fourier-based data augmentation strategy is proposed to fit the style variation of the target domain.Specifically,two images are arbitrarily selected from the source domain to obtain their amplitude spectrum and phase spectrum.The mean and variance of the amplitude spectra are matched to each other by channels to form a new data image.Secondly,to ensure that the features before and after enhancement can predict the same object and increase the discriminability between classes,an intra-domain contrastive learning method is introduced.Images of the same category in different domains are regarded as positive samples,and images of different classes in the same domain are regarded as negative samples.Different domains and different classes are excluded.Finally,the effectiveness of the proposed model is verified on Digit-DG,PACS and Office-Home datasets with diversity shift.(3)Aiming at the problems of correlation and diversity shift in the DG datasets,this paper proposes a multi-source domain generalization framework based on causal disentangled intervention,which eliminates the spurious correlation between input and labels caused by confounding factors.First,the relationship between latent variables and labels is modeled by a generative model to obtain semantic features and context-related features,and a measurement method is introduced to increase the independence of these features.Second,all contextual features are approximated by a confounder stratum,and then fused with semantic features to build an unbiased classifier.Finally,the experiment results show that the Causal Disentangled Intervention Model can achieve the state-of-the-art on the NICO dataset with both correlation and diversity shift.