Research On Multi-source Domain Adaptation Method Based On Latent Domain Information

In recent years,with the substantial improvement of computer performance and Internet big data technology,the deep neural network model has made great progress in the field of artificial intelligence.Among them,the image classification direction of transfer learning and domain adaptation has attracted extensive attention of researchers at home and abroad.Although the current performance of deep learning has been greatly improved compared with traditional machine learning classification models,there is still huge room for improvement and broad application scenarios worthy of research and exploration.(1)The current development of unsupervised domain adaptation mainly focuses on feature learning in single-source single-target scenarios and multi-source single-target scenarios,but there is a lack of corresponding research on the more common scenarios without domain label in reality.(2)In a multi-source domain scenario,there is still a data offset between each source domain and the target domain,and forced migration from less relevant domains will inhibit the performance of the target model,leading to the “negative migration” problem.In view of the above two problems,the main research contents of this thesis are as follows:In order to solve the problem of unsupervised domain adaptation in unlabeled scenarios,a fine-grained domain adaption deep network architecture for mixed latent domains is proposed to learn robust object classification by automatically discovering fine-grained information of latent domains in image datasets.device.The network framework is mainly divided into two parts.One part uses the method of combining underlying feature clustering and adversarial learning to automatically discover latent domains and divide them into several pseudo-source domains;the other part not only uses domain-specific feature extractors to separate Align the distribution of each pair of pseudo-source and target domains in multiple specific feature spaces,and combine multiple domain-specific classifiers to obtain optimal classification results.Among them,the local maximum mean difference is used in the domain-specific classifier,which makes full use of the fine-grained information of the category labels of the source domain samples,and solves the problem of aligning related sub-domain features in the latent domain adaptation problem.In order to solve the problem of data offset between multi-source domains and target domains in multi-source scenarios,an unsupervised multi-source domain adaption algorithm based on convolutional network and hybrid latent space multi-alignment is proposed.The algorithm jointly exploits domain labels,data structures,and category labels in the network,and improves domain-invariant semantic feature representation.Specifically,based on utilizing the existing domain alignment and classification result alignment,a new data structure alignment method is proposed to exploit the inherent properties of different domains.With this design,both the category consistency of the latent space can be considered,and the domain and structural differences between different source and target domains can be eliminated.In addition,in order to improve the portability of the deep neural network,the matching normalized layer is used to replace the traditional batch normalized layer to align the Latent layer of the deep neural network.Finally,the decision boundary is optimized by category alignment based on features extracted by convolutional neural network and graph convolutional neural network.The methods proposed in this thesis have been tested on public datasets,and the results of the proposed method have been compared with existing advanced research methods to demonstrate the improvement of the classification performance of the proposed method.In addition,this thesis visualizes the characteristics during the experimental process to visually demonstrate the effectiveness of the network.A large number of experiments have shown that the two network frameworks proposed in this thesis,latent domain adaptation and latent spatial multi aligned domain adaptation,respectively,provide a certain degree of solution to the problem of no domain labels and "negative transfer" in multi-source domain adaptation.