Research On Face Recognition Based On Wavelet Transform And Subspace

Face recognition is a biometric recognition technology, which bases on human face feature. It has an extensive foreground of applications in many fields, such as authentification of message, identification and human computer interaction. However, face feature always changes shape itself, and it is easy to be influenced by different illumination and view. The research of face recognition is difficult and extensible. Recently, the algebra methods of face feature extraction are popular with many researchers. And Subspace method is the most important method of them. Thus, this thesis focuses on wavelet transform and subspace method to research face recognition.Firstly, this thesis introduces the face detection method and researches mainly the method that uses Haar feature and cascade of Adaboost classifiers. The feature distance is proposed to improve selecting weak classifiers based on Adaboost algorithm. After training, the classifier is compared with the one in OpenCV library.This thesis researches the character of wavelet transform and 2D wavelet transform.2D wavelet transform is used to transform images and reduce the dimensions of features due to wavelet's multiresolution and frequency decomposing. After transform, low-frequency component keeps approximate information and high-frequency component keeps minutial information.And then, this thesis discusses two feature extraction methods based on linear subspace, which are PCA and LDA. A improved 2D-LDA is proposed based on basic LDA. The improved 2D-LDA which focuses classifying and UPCA which focuses describing are composed to realize classification and verification. The experiment is executed to compare the composed method, UPCA and the improved 2D-LDA.Finally, a design of face recognition system is proposed based on the face detection algorithm and the face recognition algorithm, which are researched by this thesis. The software system consists of pretreatment, face detection, feature transform, feature extraction and classifying strategy. This thesis proposes an face recognition algorithm which composes wavelet transform and lincar subspace. The experiment proves the algorithm has good recognition rate. The improved 2D-LDA reduces the calculating complexity during classifying and the UPCA ensures the low false match rate.