Research On Image Classification Based On Deep Domain Adaptation

Deep domain adaptation is a branch of deep transfer learning,which has emerged as a new learning technique to address the lack of massive amounts of labeled data.Compared to conventional methods,which learn shared feature subspaces or reuse important source instances with shallow representations,deep domain adaptation methods leverage deep networks to learn more transferable representations by embedding domain adaptation modules into the deep network.Deep domain adaptation has immense potential.Yet,there are still some problems to explore and study.For example,negative transfer,transfer bounds,etc.In view of the above problems,this paper proposed three solutions to improve the effect of domain adaptation and reduce negative transfer,respectively as follows:Firstly,a new network model is proposed to solve the problems of unsupervised domain adaptation,such as negative transfer and underadaptation,which appear in the existing methods of deep domain adaptation.This method combines the two techniques of feature distribution and domain adversarial learning to make them work together on the network model and improve the domain adaptability of the model.Based on the implementation of convolutional neural network,the proposed method carries out feature distribution matching in the fully connected layer of the network,leverage MMD to measure the difference of feature distribution between domains,and constructs two fully connected layers of the network after the convolutional layer as the domain discriminator for the domain confrontation.Based on a large number of experiments,this model is superior to some representative methods in classification accuracy,which proves that the proposed method can improve the domain adaptability of deep network models.Secondly,aiming at the problem of unsupervised domain adaptation,a new improvement scheme is proposed based on the shortcomings of the existing feature alignment method,which exploit entropy minimization and feature distribution alignment to enhance the domain adaptability of deep networks.The second order statistics of the source domain and the target domain are optimally aligned.Different from the current Euclidean method,the covariance matrix of the target domain of the source domain is aligned by geodesic manifold.With regard to the problem of hyper-parameters selection,a new cross validation method is proposed by introducing entropy minimization,which makes the hyper-parameters selection more efficientThirdly,a new network model is proposed to solve the problems such as the fluctuation of the model and the difficulty in selecting hyper-parameters during the training.The method is based on ResNet-50,use MK-MMD to match features of pooling layer and classification layer to realize the aim of multi-layer matching.Also added two fields discriminant,every domain use only contains A hidden layer of discriminant to reduce the complexity of the model,and make whole network light-duty,and prevent overfitting phenomenon.Finally,entropy minimization module is added to provide guidance for hyper-parameters selection.There are 21 figures,8 tables,and 119 references in this thesis.