Font Size: a A A

Pose Estimation, Recognition Algorithms And Fuse Algorithm In Face Recognition

Posted on:2010-08-30Degree:DoctorType:Dissertation
Country:ChinaCandidate:S ZhaoFull Text:PDF
GTID:1118360275455581Subject:Pattern Recognition and Intelligent Systems
Abstract/Summary:PDF Full Text Request
Automatic face recognition provides a new approach for person recognition in the modern world,and is being paid more and more attentions in recent years.manifold algorithms,discriminant algorithms,Pose estimation and feature level fusion are studied in this thesis.The following contributions are made:1.Based on traditional LLE,discriminant LLE is proposed under tow assumptions which are:a) samples can only be reconstructed with least error by samples in the same class.b) Large error will be get if samples is reconstructed by samples in other classes. DLLE is a supervised manifold learning method.In pose experiment,DLLE is proven to be very effective.In the embedded subspace found by DLLE,different manifolds can be separated,while in each manifold the underlining structure is maintained.2.LDA and its variants are summarized,based on which three new LDA variants are then proposed:CCLDA,SFisherface and SC-LDA.CCLDA is proposed base on the fact that in recognition stage correlation metric is usually more effective than L2 metric.Correlation constrains are added in objection function to guarantee the statistical independence of the projection vectors.It reached the highest performance in face expression and palm recognition experiments.Mirror symmetry is a basic feature of facial images.SFisherface firstly uses SPCA to project training samples onto symmetrical subspace,then fisher discriminant analysis is used to get the discriminant subspace.SC-LDA casts the symmetry.constrains directly onto the projection vectors, as can be seen in objection function.The experiments showed that these two method using symmetry outperformed the algorithm that directly added the mirror images into training set.In the situation the sample number is large,computational complexity of SFisherface is even lower than Fisherface.3.Canonical correlation analysis is usually used in pose estimation of a single object or data visualization.Note that face is a regular 3D object and the face union more or less has similar variation tendency in shape,CCA can be used to connect the appearance space and the pose space.KCCA is used to solve the nonlinearity in the appearance space.Finally,multi-variable linear regression is conducted to get pose parameters of a new face.Experiments showed that under range of -60 to 60 degree, average pose estimation error of KCCA is lower than LLE or Isomap,also lower than PCA subspace method.4.To overcome the shortcomings of CCA application in feature fusion,a novel supervised learning method Enhance Correlation analysis(ECA) is proposed.By redefine between class correlation and within class correlation,ECA enchants the correlation of the same class and weaken the correlation of different classes.
Keywords/Search Tags:Discriminant LLE, Symmetrical Fisherface, Correlation Constrained LDA, Symmetry Constrained LDA, pose estimation, canonical correlation analysis, CCA, feature level fusion
PDF Full Text Request
Related items