Research Of Face Recognition Based On Ensemble Learning

With the arrival of the information age,accurate user identity recognition and information security has been an important research topic.Face recognition technology,as one of biometrics recognition technologies,become a hot spot of research with its advantages of nature,concealment and convenience.In the research of face recognition,accuracy and efficiency are the goals pursued.However,due to the particularity of human face and the complexity of practical application scenarios,existing face recognition methods need further improvement in recognition performance and generalization ability.As a hot issue in machine learning,Ensemble learning aids in improving generation capacity of face recognition system and has been widely applied to address problem of feature selection,classification and regression.Therefore,we apply the ensemble learning method to face recognition problem,aiming at optimizing the generalization ability of face recognition system and increasing the recognition accuracy.The main work and contributions of this paper are listed as follows:(1)A new local descriptor named Local Mean Pattern(LMP)is proposed to address the problem that the traditional local descriptors are sensitive to the violent change of edge pixels and noise.This descriptor describes the texture information with the overall change of the gray value in the horizontal,vertical and diagonal directions.To overcome the limitation and sidedness of single feature,Symmetric Local Graph Structure(SLGS)is introduced,and Vertical-Symmetric Local Graph Structure descriptor(V-SLGS)is proposed for the lack of texture information in the vertical direction of SLGS.The complementary texture description space is constructed by combining LMP operator with SLGS and V-SLGS,then a face recognition method based on grade score fusion decision is proposed.The experimental results show that,compared with the traditional local descriptors,LMP algorithm is more robust to the edge change and noise,and has good separability.The face recognition method based on the grade score fusion decision can effectively reduce the probability of misjudgement and further improve the accuracy of the system.(2)For the reason that local mean pattern fails to render hierarchical representation according to importance of different regions of face when it is applied to face recognition,a weight computing method based on cloud model is proposed to solve this problem.This method firstly uses the sub-image method to construct classifiers for different facial regions.Then the recognition rates of different regions are input into the inverse cloud model as cloud droplet,to generate the eigenvalue which can characterize performance and stability of classifiers.Finally,we use the eigenvalues of the cloud to assign the weights,and complete the transformation from the quantitative to the qualitative.In order to better take into account the decision-making information in two aspects of the whole and the local,the overall discriminant information is further integrated after cloud weighted integrated decision,and a face recognition method with global-local dual weighted ensemble is proposed.The experimental results show that the dual weighted ensemble model effectively improves the classification performance of original LMP operator in face recognition.(3)As to the problem that face recognition method based on static weight fail to adjust the weight of classifiers adaptively,a face recognition method based on the dynamic weighted ensemble of reliability and separability is proposed after comprehensive considering the statistical performance and the distribution of the actual output components of classifiers.Compared with the static weighted ensemble method,this method can dynamically adjust the weights of classifiers according to different dynamic of the target,and take full advantages of each classifier in recognition.Experimental results on face databases show that this method can effectively restrain the interference of high deviation classifier to recognition and improve the classification performance of the system.(4)A face recognition method based on deep cooperative training and ensemble decision is proposed,aim to solve the problem of insufficient of labeled training samples in deep learning based face recognition method,as well as coarse classification mechanism.This proposed method initializes the convolutional neural network with the labeled samples,and then combines the Tri-training algorithm to update the training classifiers.In order to reduce the noise data that may be introduced in the cooperative training,the accuracy of label is improved by increasing the label reliability constraint condition.When classifying unknown samples,different weights are distributed to classifiers according to the consistency of outputs among classifiers.Finally,classification recognition is completed by the ensemble of multiple classifiers.The experimental results in the YaleB and LFW face databases demonstrate that,the proposed method can effectively reduce the dependence of the convolutional neural network model to the labeled samples,and solve the problem of insufficient labeled samples.The multi-classifier ensemble discriminant mechanism further enhances the generalization ability of the classification model.