Research On Deep Transfer Learning Method Based On Adversarial Network

In recent years,deep learning has boosted the progress of computer vision and improved state-of-the-art performance on diverse vision tasks such as image classification,object detection,and semantic segmentation.However,the remarkable efficacy of deep learning algorithms highly relies on abundant labeled training data,which requires tedious labor work on collecting labeled data or even can not achieve.This inspires scholars to leverage off-theshelf labeled data from a related domain(source domain)to improve the model for the domain of interest(target domain)by reusing and transferring the knowledge of the related domain(source domain),to solve the problem of scarcity of labeled data.The target domain may contain data collected by different sensors,from different perspectives,or under different illumination conditions compared with the source domain,leading to a large dataset shift.Adversarial networks can be used to solve this problem,reduce the differences between domains,and improve the transfer effect.Aiming at the problem of dataset shift in deep transfer learning,this paper uses adversarial learning in generative adversarial networks to align the data distribution of the auxiliary domain(source domain)and the target domain.When the target domain has no labeled data,the auxiliary domain(source domain)with a large amount of labeled data is used to complete the classification task of the target domain.This paper innovatively summarizes the research work on the adversarial domain adaptation from the algorithm application background hypothesis,proposes an adversarial domain adaptation algorithm suitable for the universal domain adaptation background,and applies it to the classification task of static facial expression recognition.The related work is summarized as follows:(1)According to the current research work on adversarial domain adaptation,the adversarial domain adaptation is defined and classified firstly.And it is divided into homogeneous adversarial domain adaptation and heterogeneous adversarial domain according to the relationship between the label sets of source and target domains.The homogeneous adversarial domain adaptation,according to the different assumptions of the algorithm on the marginal distribution,conditional distribution,and joint distribution in the two domains,is divided into four types: marginal distribution homogeneous adversarial domain adaptation,conditional distribution homogeneous adversarial domain adaptation,joint distribution homogeneous adversarial domain adaptation,and dynamic distribution homogeneous adversarial domain adaptation.The heterogeneous adversarial domain adaptation algorithm,according to the background assumption of the algorithm application,is divided into three types of heterogeneous adversarial domain adaptation: open-set adversarial domain adaptation,partial adversarial domain adaptation,and universal adversarial domain adaptation.(2)Based on the universal domain adaptation,a novel adversarial domain adaptation network Multi-weighted Adversarial Adaptation Network(MAAN)is proposed.MAAN employs generative adversarial networks to learn shared feature spaces and aligns features between the source domain and target domain through an adversarial learning process between generators and discriminators.MAAN considers that the classifier may suffer a negative transfer effect caused by non-public label spatial data in the source domain,and solves this problem through a weighting mechanism.At the same time,MAAN combines maximum mean difference and adversarial learning to solve the feature adaptation problem in the universal domain adaptation,and simplifies model threshold selection by clustering.A thorough evaluation on three benchmark datasets(Office-31,Office-Home and Digits)shows that MAAN consistently achieves higher accuracy on adaptation tasks than other adversarial domain adaptation algorithms,such as Domain-Adversarial Neural Networks(DANN)and Partial Adversarial Domain Adaptation(PADA).And MAAN is suitable for three application scenarios: closed set domain adaptation,open set domain adaptation,and local domain adaptation.(3)The proposed MAAN algorithm is applied to the classification task of static facial expression recognition,which provides an effective solution to the problems dataset shift.Experiments show that MAAN can classify facial expressions in real driving environments with the help of existing expression recognition datasets to a certain extent,and its performance is better than the convolutional neural network Res Net-50,Partial Adversarial Domain Adaptation(PADA),and Open Set Back-Propagation(OSBP).