Research On Image Classification Domain Adaptation Based On Deep Learning

Deep learning methods have greatly improved the development of artificial intelligence,but there are still many challenges in practical applications.Existing deep learning models have higher requirements for training data,not only requiring the data to have complete labeling information,but also requiring the training data to maintain an independent distribution with the input data in the actual application,otherwise the model performance will be greatly influenced.At present,the labeling of data still relies on manual labor,which means that once the application scenario or environment changes,the corresponding model needs to be retrained,and then it needs to re-acquire enough manually labeled training data.This approach is undoubtedly inefficient and detrimental to the promotion of the model.In order to solve the problems discussed above,the domain adaptation method provides a new solution.By transferring what the model learns on labeled datasets to a new dataset that lacks labeling,the model can maintain good performance on unlabeled datasets.This thesis focuses on the domain adaptation problem based on image classification,and the research work and achievements are as follows:(1)In the deep domain adaptation problem,it will cause negative transferring problem when ignoring the relationship between the different class in source and target domain.This thesis proposed a conditional distribution alignment model based on Optimal Transport domain adaptation.Based on the theory of Optimal Transportation,this method takes into account the geometric relationship between adjacent data when aligning two domains.In order to enhance the class information of features,the model uses feature fusion strategy to fuse feature information and pseudo-label information.At the same time,the model uses the dual form of Wasserstein distance to measure the distribution difference,reducing the computational cost and making the network an end-to-end model.The experiment verify the effectiveness of the proposed model.(2)For deep domain adaptation issues,redundant information in feature representation will causes poor model performance.This thesis proposed a redundant feature elimination domain adaptation model based on bi-classifier.Based on the theory of bi-classifier learning,we enhances the input data twice to obtain the features from two views,and improves the diversity of classifiers by inputting features of different perspectives into different classifiers.At the same time,by closely combining the biclassifier method and contrast learning,this thesis enables the model to capture highlevel semantic representations of the data,reducing the confusion degree between feature from different class.Finally,the proposed model recognize the samples at classification boundary correctly by aligning the label distribution.Experimental results verify that using the contrastive loss to measure the discrepancy between two classifiers can extract valid information from the data,thereby improving model performance.(3)In the bi-classifier domain adaptation model,the sample category information is not clear,and the samples at the classification boundary cannot be effectively studied.In this thesis,a domain adaptation model based on two-attribute classifiers is proposed.The differences between classifiers are defined through the correlation matrix,thereby reducing the degree of confusion between different categories of data.Using different distribution alignment losses for different classifier branches not only enhances the diversity of different classifiers,improves the complementarity of the data,but also aligns the source and target domains from different perspect.The entropy weighting is used to adjust the weights of the feature so that the feature in classification boundary can be effectively corrected.Relevant experiments verify the effectiveness of the model.