Research On Face Recognition Algorithm Based On Domain Adaptation

In recent years,face recognition which is a biometric technology with the characteristics of safety,convenience and rapidity has been applied in many fields.A good face recognition model often relies on a large number of training data.There are not enough training samples which are labeled in some special areas like public security,hospitals and so on.In this case,face recognition has the following challenges: the performance of model trained on the source domain degenerates obviously when it is applied to the target domain;if only the gallery samples are studied,within-class scatter matrix of the gallery will be reduced to zero,and most of the discriminant analysis methods cannot be applied,the model is easy to be under fitted and has poor generalization ability.Domain adaptive methods can achieve the purpose of label prediction by transferring the discrimination information of the supervised source domain to the unsupervised target domain.But most of the domain adaptation methods proposed before cannot fully obtain the diversity of samples,which lead to the lack of diversity of the sample features and the discrimination model is worse.In addition,because the target domain is not labeled,it is unable to train the target domain samples directly.By being constructed the source domain data,the data distribution of source domain can approach the target domain,but it still cannot reach the same distribution as target domain.Therefore,how to learn more abundant and accurate discrimination information is a challenging task.Aiming at solving above problems,in detail,the contribution of this dissertation are as follows:(1)A multi-subspace face recognition framework based on domain adaptation is proposed.The innovations of this study are as follows: 1)In order to make full use of the sample information of multiple source domains,a multi-source domain face recognition method based on domain adaptation is proposed,and introducing the theoretical derivation;2)the feature representation of samples in a single subspace is unique,in order to preserve the diversity of sample features,a multi-subspace face recognition framework based on domain adaptive is designed.The same sample can retain different discriminant features in different feature spaces,and it can increase the diversity of sample features,so the performance of the model improves.In addition,we analyze its operability and advantages theoretically.(2)A face recognition method by using label reconstruction based on domain adaptation is proposed.The innovations in this dissertation are as follows: 1)Learning a similarity matrix and a prediction label probability matrix for unsupervised target domain,and converting a linear programming problem to obtain the best prediction label probability matrix,then,obtaining the prediction labels of target domain samples;2)learning a discrimination model for target domain data by using a method similar to linear discriminant analysis.In the model learning,the target domain data is used to estimate the intra-class scatter matrix of the gallery,and then the discrimination information of the gallery is added to improve the performance of the model.A large number of experiments are carried out in single data set and cross data set respectively,and a large number of experiments show that the algorithm has good performance.