Image Classification Algorithm Method Based On Unsupervised Domain Adaptation

It is expensive to have a large amount of labeled data in the field of machine vision.In the traditional supervised learning task,when the labeled data is insufficient,the learned model does not have good generalization power and cannot meet the requirements of reality.In this context,domain adaptation is proposed as a new technology.In this way,a classifier learned from a source domain with rich label data is used for a target domain with only a small amount of data with labels or no labels at all,and the knowledge learned from the source domain is used for targets without labels.The method in the domain is called unsupervised domain adaptation.At present,the main shortcomings of the unsupervised domain adaptation method are:1)due to the existence of domain shift,the samples in the domain class that are far away from their class center are highly misclassified;2)the existing domain adaptation research often ignores the implicit label data of the target domain,which leads to poor performance of the learned model.3)in the existing research,there is no deeper consideration of the similarities and differences between the inter-domain classes and the intra-domain classes.Therefore,in order to solve the above problems,the main research of this paper is as follows:First,in order to solve the problem of higher error rate of samples at the edge of the domain class due to the shift between the source and target domains,a domain adaptation method Adversarial Domain Alignment Feature Similarity Enhancement Learning for Unsupervised Domain Adaptation(AASE)is proposed.It learns domain invariant features by adversarial game and correlation alignment to reduce the domain gap,and makes these features having better discrimination via joint central discrimination and feature similarity enhancement.AASE makes the learned features have better intra-class compactness and interclass separability.To better implement the proposed function,an adversarial learning network with dual-channel parameter sharing is designed.In addition,the network parameters are learned through the gradient descent method,and the output of the classifier is used to realize the prediction labels of the target domain data.Secondly,in order to solve the situation that the existing methods do not show the minimization of the domain shift because of ignoring the class label information of the target domain,a method of deep hash domain adaptation based on adversarial learning domain class discriminative learning is proposed.First,the structure of a deep hash domain adaptive network based on adversarial learning is designed,which consists of an extractor that does not share parameters,a domain classifier,and a hash encoder that shares parameters;then,in order to use the class label information of the source domain and the target domain,the Adversarial Domain Class Discrimination Learning Deep Hash Domain Adaptation(ADHDA)method is proposed,and the domain class identification loss function is designed.Its function is to combine the display label information of the source domain with the implicit label information of the target domain,so as to realize the joint optimization of the intra class and inter class differences,and improve the performance of the domain adaptation;finally,the network parameters are learned by the gradient descent method,and the output of the classifier is used to predict the label of the target domain data.Third,to fully realize the similarities and differences between domain classes and within domain classes,and achieve the performance of a more effective domain adaptation method,a domain adaptation method based on hash coding for nuclear domain difference identification is proposed.1)The entire method is implemented based on the structure of the network adapted to the deep hash of adversarial learning;2)the Hash Coding Kernel Domain Differential Discrimination for Domain Adaptation(HKD~3A)method is proposed,and the kernel domain discrpancy discrimination loss function is designed,which maps the extracted features to high-dimensional space by kernel function,and combines the loss function of class label information in target domain and source domain and the kernel function is used to calculate the domain shift between domain classes and between domain classes;3)the gradient descent method is used to learn the parameters of the network,and the output of the classifier is more useful to realize the label prediction of the target domain data.The AASE,ADHDA,and HKD~3A methods mentioned in this article are fully tested on the digital recognition,office-home,and office-31 datasets,and compared with existing domain adaptation methods.It is concluded from the analysis that the AASE,ADHDA and HKD~3A models mentioned in this paper provide more effective method models for unsupervised domain adaptation.