Face Recognition Algorithm Research Based On Weighted Wavelet Decomposition And Fisherfaces

With the development of social economy and the progress of computertechnology, people’s request for the concealment and security of individualauthentication information is rising. And biometric identification technology usinghuman biological characteristics for personal identification has gotten a highrecognition in the field of security verification. Including, face recognition technologyhas extremely broad application prospects for its good performance of easy to use,high precision, good stability, high cost performance and difficult to counterfeit.Based on the research background and current situation of the development，thepaper is summarized systematically, in which, two kinds of methods of geometricfeatures basis and template basis in face recognition are introduced. And the contentof image preprocessing, the weighted wavelet decomposition, Fisherfaces method andthe classifiers of nearest neighbor and SVM, etc, are also deeply analyzed. In addition,the face recognition system visual interface is designed and completed by GUIDEdevelopment platform. The main work of the paper is around the following aspects:Firstly, in order to balance the energy spectrum of the image, the whitening isadded to the face image preprocessing. And for the problem that the binarization leadsto information loss, the binarization process is removed to optimize the image quality.Different from the traditional2D wavelet transform merely with the low frequencycomponent after the transform, the weighted wavelet decomposition is presented inthe paper. That is to say, the weighted combination is with the low frequencycomponent, horizontal component and vertical component, and the diagonalcomponent is abandoned with more interference information. In addition, the bestwavelet basis, decomposition level and weight coefficient group of the weightedwavelet decomposition are selected by contrast experiment, in order to reducedimensions effectively and retain more information for classification.Secondly, Fisherfaces that combines PCA and LDA algorithm is used to solvethe problem that scattering matrix is singular within class and PCA is not conduciveto sample classification in dimensionality reduction process. For different sample size,the best normalized dimensions and characteristic dimension are determined. Compared with the traditional method of face recognition, the recognition effect ofthe method presented in the paper is markedly improved, and the recognition rates of98.75%and100%are achieved respectively in experiments on ORL and YALEdatabases. In addition, this method has good robustness against the noise.Thirdly, the nearest neighbor classifier and SVM are studied and programmablyimplemented. Improved SVM is applied to the face classification, using Libsvmtoolbox, and substituting the grid search method for the traditional parameter iterationmethod, to greatly reduce the computational complexity.Fourthly, according to the flow diagram of the face recognition system, a simpleface recognition system interface is designed and completed by GUIDE developmentplatform which MATLAB provides. And it can be applied to two datebases of ORLand YALE and two classifiers of nearest neighbor and SVM. The functions of imagepreprocessing, the weighted wavelet decomposition, the establishment of the trainingand testing databases, feature extraction with Fisherfaces as well as face decision, etc,are also implemented.