Research On Face Recognition Based On Fusion Method

Face recognition is one of the hot research fields in biometrics recognition.With the development of machine learning and deep convolutional neural networks,various traditional face features and feature extraction methods for face recognition have been diversified.The problem of face recognition is the change of attitude,expression and illumination.Traditional facial features are relatively single and perform poorly under large posture and expression changes.The deep networks use the face image with good illumination for training,which has a bad effect on the drastic change of illumination.This paper studies typical traditional Gabor features and typical depth features and performs feature fusion at different levels.The proposed face recognition methods are as follows:Firstly,for the problem that single Gabor face feature has poor effect under the changes of posture,expression and illumination,this paper proposes the face recognition method based on Gabor feature and the fusion of different kernel functions.The Gabor features of different scales and directions,extracted by two-dimensional Gabor wavelet,are fused in the early fusion.In the late fusion,the rbf and linear kernel functions of support vector machine are fused and the parameters between the kernel functions are determined by KMOD algorithm.The final experimental results show that the fusion method can improve the traditional face recognition in machine learning.Secondly,in view of the fact that the fusion of complementary depth features can improve the performance of single feature,this paper designs a face recognition method based on the sparse representation of Facenet and the integration of convolutional network.Via combining the convolutional feature of Facenet with the sparse representation,the sparsely represented face recognition method based on Facenet is proposed.The normalization dictionary is constructed by convolutional features of Facenet to realize sparse representation classification.The loss function,combing softmax loss and center loss,is retrained on the pre-training model of Facenet to compare the effects of face verification and recognition.The hard voting strategy is used to integrate the sparsely represented face recognition method based on Facenet and different convolutional networks to improve the face recognition effect.The final experimental results show that the ensemble method can further improve the recognition rate on the basis of each convolutional network.Finally,aiming at the problem that the deep neural network has poor performance against the drastic illumination,this paper proposes the face recognition method combining GIR feature and convolutional feature.The GIR model is used for the illumination processing of the face image to obtain GIR features.According to the different local areas in the illumination processing process,GIR features can be divided into EGIR-face and BGIR-face features.Through distance measures,GIR features are respectively fused with the convolutional features of Facenet,Insigtface and Resnet50,and face recognition is achieved through nearest neighbor analysis.The convolutional network can achieve the advantage of GIR model and pay more attention to the influence factor of illumination in face recognition.The final experimental results show that the face recognition method combining GIR features and convolutional features can indeed improve the face recognition effect under severe illumination changes.The datasets used in this paper are ORL,Ext Yale B,and LFW datasets.Based on the results of comparative experiments,it can be found that the methods proposed in this paper do improve the face recognition effect.