Cross-Domain Image Classification Based On Domain Invariant Feature Representation Learning

Recent studies have shown that the image classification model is often affected by the domain-gap between the training set and the test set,resulting in a decrease in its performance in practical applications.Therefore,the research on cross-domain image classification methods has received many attention.The common domain-gap is generally affected by the diversity of color styles and art styles,resulting in the differences between the domain distribution.This domain-gap is visible to the human eye and is an obvious difference at the image level.Another special kind of domain-gap is the subtle difference at the pixel level.One kind of this domain-gap is the distribution-gap between the adversarial example domain and the clean image domain.The domain-gap between two different domains is very subtle,but this domain-gap will be magnified in the classification model and seriously affect the classification results.From the perspective of domain invariant feature representation learning,this paper proposes two methods to improve the generalization ability of cross-domain image classifier in two different scenarios.Aiming at the problem that domain-invariant features are difficult to construct directly,we propose a method based on domain-specific feature ablation.We indirectly construct the invariant features of the domain by explicitly modeling the specific features of the domain and performing feature ablation,which solves the adaptive problem of multi-source domain in the scene where the domain differences are visible,and improves the generalization ability of the cross-domain image classification model in the target domain.We conducts experiments on multiple public datasets and the experimental results prove the effectiveness of the proposed method.Aiming at the problem of cross-domain classification in which domain-gap is subtle,this paper proposes a method based on zero-knowledge image transfer cycle to reconstruct the features in two domains into robust domain invariant features.This method solves the dependence on the distribution information of adversarial example domain,and improves the accuracy of the model.This paper conducts experiments on multiple public datasets and the experimental results show that the proposed method is effective in improving the model generalization ability.