Domain Generalization In Deep Learning-Based Image Classification

Deep Learning has achieved remarkable success in various areas,especially in computer vision area.Classic deep learning methods are built on the i.i.d.assumption that training and testing data are independent and identically distributed.However,in real scenarios,the i.i.d.assumption can hardly be satisfied,which lead to the sharp drop of classic deep learning algorithms’ performances.So the key to this problem is how to generalize deep learning models to unseen data distributions,which is called domain generalization(DG),i.e.,out-of-distribution generalization.Domain generalization deals with a challenging setting where one or several different but related domain(s)are given for training,and the goal is to learn a model that can generalize to an unseen test domain.Here domain means the data distribution.Great progress has been made in the area of domain generalization for recent years,the main idea of which is based on representation learning.Many methods of representation learning are based on adversarial training or feature disentanglement to extract domain-invariant features,which can improve model’s generalization ability.However,most methods based on adversarial training only match the marginal distributions of data which lack considerations for label information;while methods based on feature disentanglement lack efficient training strategy to improve the performance of disentanglement.Based on these situations,we thereby do a research of current mainstream research methods and analyze their advantages and disadvantages,research and solve their problems.Then we propose a model with better domain generalization performance,Adversarial Domain-Invariant Variational Autoencoder(ADIVA).The main work and contributions of this paper are as follows:(1).Based on the research and analysis of current mainstream research methods,we combine classic models in methods based on feature disentanglement and adversarial training and build the model framework of ADIVA,then we prove its identifiability,which provides solid theoretical foundations.(2).To solve the problem of methods based on adversarial training,based on the causal relationships in the domain generalization task,we propose distribution contrasive loss and weighted classification loss to match the corresponding distributions,which solve the domain shift problems.To solve the problem that methods based on feature disentanglement lack efficient training strategy,we add adversarial training module in ADIVA.Adversarial training can remove the confusing information of other factors for factors obtained by feature disentanglement,which improve the effect of feature disentanglement.(3).We do domain generalization experiments with ADIVA model on two benchmark datasets.Through the analysis of results and comparison with other domain generalization methods,it shows the effectiveness and superiority of ADIVA in domain generalization problems.