| Face recognition technology is a kind of biometric identification technology,because of the data collection friendly,the objectivity of the face and the diversity of application scenarios.It has become a hotspot in pattern recognition and in-depth learning.But face recognition in the specific application process will encounter a variety of practical problems: First,face images are susceptible to external factors such as light,gestures,makeup and so on;Second,the feature extraction of face images,different feature extraction methods for the final recognition has a pivotal role.The traditional image feature extraction is generally carried out from the bottom aspect of the texture,shape,color,etc.,it is difficult to extract facial images of nature structural information.The development of manifold learning theory provides a new idea for the feature extraction of high-dimensional data,and the related research shows that face data is more likely to be distributed in high-dimensional nonlinear manifolds,so nonlinear dimensionality and manifold learning,therefore,the application of nonlinear dimensionality reduction and manifold learning theory in image recognition,especially face recognition,has been paid more and more attention.Based on manifold learning,this paper mainly studies the local linear embedding algorithm and the supervised local linear embedding algorithm.Because of that the sample points that deviate from the overall distribution of the samples may be mapped to other planes in the low-dimensional reconstruction process and combined with the advantages of Kmeans ++ clustering algorithm,a CSLLE manifold learning algorithm based on clustering is proposed.At the same time,the CSLLE algorithm introduces new parameters,lack of objectivity and the poorness of the noise between the class and the class spacing,The distance similarity measure in the algorithm improved,the generalization ability of the algorithm is improved,and the result is also obtained in the relevant data set experiment.The main research work is as follows:1.Based on the global and local dimension reduction methods,such as principal component analysis,multidimensional scale analysis,Laplace feature mapping,etc.are discussed in detail in manifold learning.And the comparison and analysis of the algorithms on the different data sets are studied,and the advantages and disadvantages of each algorithm in different data sets are analyzed.2.On the basis of manifold learning,the LLE algorithm,the SLLE algorithm that references the sample category information and the value of the parameter in the specific application process are analyzed in detail.The SLLE algorithm uses the category labels of the samples to measure the similarity between the data points,but ignores the effect of the individuals with large differences in the data set on the overall data,Thus the Kmeans ++ clustering algorithm is used to identify the "anomaly point",and the distance matrix between the data points is further improved to improve the generalization ability of the algorithm.The Yale and ORL face data sets show the feasibility and generalization of the algorithm.3.In the SLLE and CSLLE algorithms,the similarity between the distance between the interclass classes and the distance between the classes is linear,so that the discrimination and generalization ability of the embedded data are still limited to a certain range.And the presence of noise in the sample will destroy the neighborhood relationship between the samples,affecting the results of the experiment;In addition,the CSLLE algorithm improves the recognition rate,but also introduces new uncertainties: The problem of the new parameter increases the subjectivity of the algorithm.In view of this situation,under the inspiration of the original algorithm,the measurement of the distance between the classes in the class is changed,the number of parameters is reduced and the noise is also constant Inhibitory effect,which contributes to the low-dimensional embedding representation of face data. |
PDF Full Text Request