Research On Deep Transfer Learning Cross-domain Image Classification Algorith

The traditional supervised learning and semi supervised learning algorithms assume that the training and test data follow the same distribution,it is difficult to meet in real situations,but transfer learning solves the problem.By using data from similar domains with knowledge learned in relevant domains,the time required to obtain samples of new labels is reduced,and learning efficiency is greatly improved.Based on the image classification algorithm of transfer learning,this paper proposes a new transfer learning model algorithm for the problems of domain shift,domain gap,and feature loss in the transfer process:Cycle Selection Pseudo-label Classification model(CSPL)and Double-weighted Universal Domain Adaptation model(DW-UDA).Unsupervised domain adaptation,as the main research focus of transfer learning,it is also an important method for its implementation and the mainstream method at this stage is to match the probability distribution of the target domain and the source domain as well as the idea of adversarial based knowledge transfer.However,while the existing methods match the probability distribution,the confidence degree of the pseudo-label and the loss calculation of the pseudo-label are not considered in the selection of pseudo-labels.Therefore,aiming at the adaptive problem in the unsupervised field,A cyclic threshold selection pseudo-labeling method is proposed and add the pseudo-label loss function for backpropagation training after fully considering the confidence degree of pseudo-labels.Experiments on the commonly used domain adaptive datasets Office31,Office-Home,Image CLEF-DA and Amazon-Review show that the model improves the transfer accuracy by about 5% on average compared with the traditional algorithm and the current domain adaptation model.Among them,the small dataset Office31 increased by about 1.2%,the larger dataset Office-home increased from 70%to more than 80% compared with the traditional model.There are also different degrees of improvement on other datasets,among which the accuracy of the existing model on Image CLEF-DA has increased from about 77% to more than 81%,and the average improvement is 2%～8% compared with traditional and current domain adaptation models on Amazon-Review.As an in-depth stage of the development of domain adaptation,universal domain adaptation includes the research situation of all domain adaptation: its target domain and source domain contain both public and private classes.So how to reject private classes and find public classes is the main way to do knowledge transfer.Most of the existing general-purpose domain adaptation model algorithms solve the domain displacement and domain gap by weighted loss calculation,but with the increase of samples and the inability to improve the distribution of samples under unsupervision,a single weighting algorithm cannot well express the specific situation of the target sample and cannot effectively carry out knowledge transfer.Therefore,this paper proposes the DW-UDA model,which fully considers the characteristic spatial information of the sample while ensuring the transfer efficiency by maximizing the pseudo-margin and the spatial distance information between the target sample and the source domain as the weighting of the source domain sample,and effectively improves the accuracy of knowledge transfer by understanding the target sample more comprehensively.In the benchmark domain adaptive data set test and verification model,the average transfer accuracy rate on the Office31 dataset is 1.32%,the Officehome medium high appearance model is about 2%,and the large dataset Image NetCaltech and Visda2017 are all improved by about 3.4%.In addition,experiments show that the model can effectively reduce the impact of negative transfer.This paper also verifies whether the algorithm theory is in line with reality through various evaluation indicators,including transfer accuracy,feature space visualization,ablation experiment and negative transfer,etc.,and compares the experiment with the current mainstream transfer algorithm,and the results reflect the advantages of the algorithm proposed in this paper,which can effectively complete the transfer tasks.At the same time,there are still deficiencies in the consideration of negative transfer and the complex process of transfer,such as efficiency,timeliness,personalization,etc.,which is also the focus of further research in the future.