Study Of Facial Expression Images Recognition Methods Based On Manifold Learning

There is much information hidden in face expression and it will contribute to exploring personal emotion and promoting its application in human-computer interaction and security monitoring by face expression images identification.Thus,in this thesis,the study background and status of face expression images recognition are firstly discussed,then two feature extraction methods based on manifold learning are presented to classify face expression images efficiently.The main works of the thesis are stated below:1)A maximum nonparametric margin projection(MNMP)method is put forward.In the proposed MNMP,in order to overcome the problem that linear discriminant analysis(LDA)cannot handle non-Gaussian distribution data well,a between-class scatter and a within-class scatter are nonparametric or locally defined.Moreover,a margin metric is constructed to avoid small sample size problem.Finally,an objective function can be modeled to detect a linear subspace with the maximum margin criterion.Experiments on benchmark face expression images show that the proposed MNMP can extract features from face expression images with high efficiency.2)A distance weighted locality preserving projection(DWLPP)is brought forward.In DWLPP,the class information of face expression images is introduced to model a distance-weighted matrix,by which the diversity between intra-class images and inter-class images can be adjusted.And then,LPP is adopted to project these distance modified face expression image data to a low dimensional discriminant subspace.Experimental results on some face expression images validate that DWLPP is superior to other manifold learning based methods.