The Research Of Face Recognition Based On Transfer Learning And Feature Fusion

At present,the traditional identification method is not easy to carry,' and is easy to loss,damage and sometimes at the risk of cracking or steal.Thus,face recognition has received the widespread attention in the field of biometrics,because of its stability,strong concealment and individual differences,it can guarantee the security,and face recognition has more and more widely application fields,such as security,civil,military and other fields.Face recognition also has very important significance in development of image processing field.In practical applications,face recognition will encounter the problem of small sample dataset,in other words,the numbers of the training samples are far less than the dimension of the test face sample,traditional methods of feature extraction and classification in face recognition will hardly obtain strong robust and good recognition accuracy in the problem of small sample dataset.This paper is aim to solve the theme of how to efficiently extract face feature and realize face recognition in the small sample dataset,so as to improve the robustness and accuracy in face recognition.This paper study and realize the face recognition method based on transfer learning and feature fusion,the main study and innovation points include:Firstly,considering the problem of similarity between-class,difference in within-class,we propose a transfer learning method based on sparse representation.This method combines sparse representation and transfer learning.Test face sample can be represented as a linear combination of all original training samples,and then we can compute this test face sample's sparse matrix.We define label samples which include label training samples and label test sample in order to realize completely same within-class and completely different between-class,and realize label test sample to have the same distance to each class.We transfer the previous sparse matrix to the label training samples so as to generate the reconstructed label training samples which can exploit the discriminative information hidden in the different sparse coefficient vector.The reconstructed label training samples can obtain the reconstructed errors of label test samples by the sparse representation model,and these reconstructed errors can be used to determine the class.Secondly,considering the problem of low classification credibility,by fusing original samples' reconstructed errors and label samples' reconstructed errors,we propose a classification method based on weighted fusion scheme.The original training samples can obtain the test face samples' reconstructed errors by sparse representation model,and the reconstructed label training samples also can obtain the label test samples' reconstructed errors by sparse representation model.After the experiment,the average Pearson correlation between these two kinds of reconstructed errors is very low,which indicates that original training samples and reconstructed label training samples are complementary representation for the same class.Thus,we fuse the weights from those two different reconstruction errors and finally classify test face sample by evaluating which class leads to the minimum weighted fusion.Thirdly,in order to address the problem of sparse representation of only considering the sparsity between samples but not considering the constructive information in samples,we propose a method of face recognition based on feature fusion.This method fuses the sparsity between samples and the constructive information of average hash feature in samples,we calculate test face sample's sparse matrix on original training samples by sparse representation model.In the same time,we calculate average hash features of all the original training samples and test face sample,and finally reconstruct samples by exploiting the previous sparse matrix and average hash features of original training samples.The reconstructed errors between reconstructed samples and average hash feature of test face sample are used to determine the class which test face sample belong to.This method can improve the whole dataset's recognition accuracy and robustness.