| Face recognition technology has a wide range of application prospects,so it has attracted the attention of researchers.Due to the high dimension of the sample,it is necessary to reduce the dimension of the sample in order to achieve better recognition effect.Dimension reduction is also called feature extraction.For face recognition,it is to extract low-dimensional features conducive to classification,so as to obtain good classification results.There have been many algorithms.Among them,the feature extraction algorithm based on graph embedding can extract the internal manifold structure information from the samples.It has been widely used in many fields,such as image classification,pattern recognition and so on.However,such algorithms are sensitive to the number of nearest neighbors,occlusion and noise.Therefore,this paper deeply studies the face recognition algorithm based on graph embedding.The main research work includes two parts:（1）Based on the two PFLPP algorithms of SLSDP,aiming at the shortcomings of the two algorithms,adaptive median neighbor and diversity preserving algorithm（AMNDPA）is proposed.According to the sample category,median intra class distance and median inter class distance,AMNDPA gives a new adaptive nearest neighbor and difference weight calculation method.The weight calculation of the sample adopts the cosine distance of the sample.For the determination of the nearest neighbor of the sample,an adaptive configuration method based on median intra class distance and median inter class distance is given.At the same time,in order to further enhance the performance of the algorithm,Orthogonal method was used.The improved algorithm not only improves the recognition rate,but also has certain noise resistance.（2）In order to make the algorithm have the ability of feature subspace learning and feature selection at the same time,and further improve the recognition performance and noise resistance of the algorithm,based on boundary discriminant analysis,Parameter-Free marginal discriminant analysis based on L2,1-norm regularization(PFMDA/L2,1)is proposed.PFMDA/L2,1 presents a method for calculating the weight of nonparametric nearest neighbor samples and determining the adaptive nearest neighbor.PFMDA/L2,1 uses the cosine distance of the sample as the weight.For similar samples,due to the small number of samples,the processing method that the same kind is the nearest neighbor is adopted;For heterogeneous nearest neighbors,it is adaptively determined according to whether the cosine distance of the two sample pairs is greater than the cosine distance of the two types of mean,so the algorithm does not need any parameter setting.At the same time,L2,1 norm regular term is added to the objective function.Through the constraint of the regular term,the rows of the projection matrix can be sparse,which ensures that the obtained characteristic matrix has the functions of feature selection and subspace learning.Experimental results show that the method is effective and has good noise resistance.Figure[17]Table[15]Reference[65]... |
