Research On Single Sample Face Recognition Algorithm Based On Domain Adaptation

Many face recognition models depend on the scale of training data.In many practical applications,such as,face unlock,face payment,etc.,there is only one sample in gallery for recognizing which is often called Single Sample Face Recognition(SSFR).In this situation,face recognition will face with great challenges,for example,each person's intra-class scatter matrix will degenerates to 0 which makes most discriminant analysis methods unable to work and the trained model is easy to underfit and has poor generalization ability.Thus,it is very challenging to improve the accuracy of face recognition!At present,there are many effective algorithms proposed for SSFR.From the angle of gallery,SSFR algorithms can be divided into gallery sensitive and gallery insensitive.From the angle of samples,SSFR can be divided into three categories: unsupervised learning,additional sample generation and learning methods base on the auxiliary datasets.These methods work well while faced with the problem of single sample face recognition,but there still are some problems,for example,supervised knowledge cannot be used in gallery to guide model learning in unsupervised methods and it is hard to guarantee the quality of samples generated by additional sample generation methods,besides,the intra-class variations of data in Gallery are difficult to be estimate by the methods based on the auxiliary datasets when the distribution of auxiliary dataset is different from that of gallery.Single source transfer learning effectively solves last problem.In reality,there may be more than one supervised source domains,but the current methods cannot fully utilize the knowledge of multiple source domains.To solve the problem mentioned above,innovative work of this paper summarized as follows:(1)In order to make full use of the knowledge of multiple source domains to help the target domain to learn a better model,we transfer the knowledge of multiple source domains to the target domain to learn a Gallery insensitive model,besides,we propose a method of SSFR based on multi-source domain adaptation: targetize multi-source domain(TMSD),which could transform images from multiple source domains into target domains.We use dataset Multi-PIE and CAS-PEAL-R1 to test the accuracy rate,the experimental data verified the effectiveness of TMSD.(2)Due to that some close relationship exists between the gallery and the testing set.In order to improve the discriminability of the model for probe,we use the samples in the gallery to guide the learning of the model and present a gallery sensitive single sample face recognition algorithm based on field adaptation: discriminative domain adaptation(DDA).In addition,seven different single-sample face recognition algorithms are used to carry out comparative experiments in five data sets,which are MULTI-PIE,CAS-PEAL-R1,OFD,AR and FERET.Experimental results demonstrate the effectiveness of DDA.The innovations in this article are as follows: 1.Propose a method of SSFR based on multi-source domain adaptation: TMSD,and provide specific optimization solutions for TMSD;2.A method for estimating the scatter matrix within the gallery sample class based on the common subspace is proposed.3.Give a detailed description of the way how to use the target intra-class divergence matrix of the source domain and the coefficients learned from the common subspace to infer the Gallery's intra-class in the original space.4.Learn a feature extractor by discriminant analysis method which uses four elements including: the inter-class scatter matrix of the targetize source domain,the total scatter matrix of the target domain,the inter-class scatter matrix and the total scatter matrix of the gallery.