Robust Face Authentication using ESD & Optical Flows in LDA SpacePresentation OutlineProblem FormulationProposed ApproachOptical Flow ResidueIndividual Eigenspace DecompositionLinear Discriminant Analysis (LDA)LDA Space (Continued…)Implementation & ResultsReferencesRobust Face Authentication using ESD & Optical Flows in LDA SpaceAhtasham Ashraf901 749 3157ECE 539 Project Fall 2001Presentation Outline•Problem Formulation•Proposed Approach•Implementation•Results•ReferencesProblem Formulation•The main idea is to authenticate different Subjects in the presence of facial expression variation & registration errors along with illumination differences.Proposed Approach•Traditional appearance based methods (simple PCA) when applied directly to image pixels are very sensitive to shifts, rotations, scale & expression variations.•Optical Flow residues are commonly used in analysis of human expression recognition.•I have used Individual Eigenspace Decomposition along with Optical Flow Residues in Linear(Fisher) Discriminant Analysis for subject classification.Optical Flow ResidueOptical Flow between two images indicates the amount of motion between the two frames. The images of a same person with expression variation or registration errors will show optical flow as:Optical flow ResiduePredicted imageOptical FlowTwo Images for OPFIndividual Eigenspace Decomposition•Universal Eigenspace decomposition : A single space shows inter & intra subject variation.•But what we want is robustness to expression & illumination variation within a single subject.•This suggests a potential metric for face authentication: “Individual Eigenspace Decomposition”•The residue between an image & its reconstructed eigenspace image is called “The EigenSpace Residue”.Individual EigenSpace Residues for one SubjectSubjectLinear Discriminant Analysis (LDA)•LDA is used to construct low dimension features from a high dimension feature space. It can be applied to two or more classes.•We can make two classes for each subject: “Self class” & “Imposter class”.•Both Optical Flow Residue & Eigenspace Residue have the ability to discriminate between two classes. We can combine these two using LDA to get an LDA space for “each” Subject.•So we have Optical flow residue on one axis & Eigenspace residue on the other.LDA Space (Continued…)Implementation & Results•The coding for this project is done in Matlab.•The image data base was acquired from Yale & Olivetti Research Laboratory.•I have used Block based optical flow calculation method by “Lucas-Kanade”•For Individual Eigenspace Decomposition , I have used Turk and Pentland's method.• Its giving very good authentication results inspite of expression variations, registration errors & Illumination differences.Results:References•Turk, M. and Pentland, A. (1991) "Eigenfaces for recognition". Journal of Cognitive Neuroscience, 3( 1), pp. •Young, D. (1994) First Order Optic Flow•Duda "Pattern Classication", Richard O. Duda, Peter E. Hart and David G. Stork, Wiley-Interscience,2001.•Moghaddam, B. and Pentland, A. (1997). Probabilistic Visual Learning for Object Representation. IEEE Transactions on Pattern Analysis and Machine Intelligence. Vol. 19, No. 7.
View Full Document