Research On Source-free Domain Adaptation Methods Based On Deep Learning

With the rapid development of deep learning,the field of computer vision has achieved unprecedented success.As a basic task in computer vision,image recognition has also made rapid progress,playing an important role in industrial production and people’s daily life,greatly promoting industrial development and providing convenience for people’s daily life.The existing image recognition methods have achieved good performance in an ideal test environment.However,in the more complex and changeable actual scene,when the source domain model is directly applied to the target domain,the performance will be significantly reduced due to the inconsistent data distribution of the two domains,such as the inconsistent style of the image,the impact of noise in the image,and so on.Based on this,researchers put forward the concept of domain adaptation,aiming at migrating the source domain model to the target domain,so that it can work well.Traditional domain adaptation methods require access to the source domain data during the adaptation process,but this is unrealistic in some specific scenarios,such as data privacy,data transmission and storage restrictions.Based on the above challenges,this thesis mainly focuses on source-free domain adaptation,that is,only using the trained source domain model to achieve domain adaptation without accessing the source domain data.This thesis first summarizes the existing source-free domain adaptation methods,and then proposes corresponding solutions to the problems faced by the existing methods.The main work of this thesis is summarized as follows:1.This thesis proposes a source-free domain adaptation method based on pseudolabeling technology.First,the entropy regularization is used for initial feature alignment.Then,in order to solve the problem of low quality of pseudo labels in many methods,a pseudo-labeling method based on confidence and clustering is proposed to obtain more accurate semantic information and correct the deviation caused by entropy regularization.Finally,the idea of mutual learning is applied to label denoising to improve the expression ability of the model.2.This thesis proposes a source-free domain adaptation method for generating auxiliary source domain.First,select a batch of samples similar to the source domain in the target domain as a bridge between the two domains,and then treat it as a semi-supervised learning problem.Use mixup training to align the source domain,and then indirectly achieve the feature alignment between the source domain and the target domain.3.This thesis proposes a source-free domain adaptation method based on knowledge transferability estimation.At present,most methods do not consider the transferability of various knowledge in the source domain.This method uses uncertain distance as a tool to quantify knowledge transferability,and proposes a knowledge transferability estimation module,which can measure the transferability of knowledge and the reliability of semantic information in the target domain.Then an adaptive gating mechanism is proposed to select knowledge and screen the semantic information of the target domain,which solves the two main problems faced by the existing methods.