| Face Recognition is one of the important issues of Pattern Recognition. Butthe images of human face are non-stationary signals, the features for classificationare usually in the local time-frequency information. It is di?cult to extract e?cientfeatures by common transform methods. Wavelet transform, which has been de-veloped recently, is an e?cient method for analyzing non-stationary signals, and itcan extract the multi-scale features just like the human visual sense. So wavelet hasbeen used in face recognition.Wavelet packet transform, which can provide an ar-bitrary time-frequency decomposition for the signals, contain more wavelet packetbases for classification.The paper has used the classical wavelet packet Local Discriminant Basismethod to Face Recognition.And has made a new Face Recognition method withcombining Princial Component Analysis method.1.Wavelet PacketThe obvious defect of traditional Fourier Transform is no local information,and it is only fit for the known and steady signal. Wavelet Transform overcomesthe defect of Fourier Transform which has bad time-frequency local character, andovercomes the defect of Short Time Fourier Transform which has the fixed reso-lution and hasn't fast algorithm. It is adaptive to the time-varying and stochasticsignal. The character of Wavelet Transform is fit for seismic signal. Though theresolution of Wavelet Transform isn't fixed in time-frequency field, its frequencyresolution will reduce as frequency's increasing. So it's not fit for practical de-mand, the resolution should be adjustive randomly according to the character ofsignal and demand. Therefor, wavelet packet theory is producing. Wavelet packetanalyzing is a more subtle decomposing method and has orthogonal character. Ithas the adaptive capability according to the signal, so wavelet packet decomposingis more fit for analyzing signal in time-frequency field.Being developed from wavelet transform,wavelet packet transform is also atime-scale and time-frequency signal analysis method, which provides a group of orthonormal wavelet packet basis with perfect localization property both in timedomain and frequency domain. Being a signal processing tool,wavelet packet trans-form has many desirable Properties and is widely applied to many fields.2.Princial Component AnalysisPCA is an e?ective method of data analysis in statistics.The base of the the-ory is K ? L translation.The aim of the PCA is finding the best vector that couldreplace the larger one.It could make the data space from m to n(n m) and savethe primary information.The recognition space that gotten from PCA is the bestapproximation to the original vector.This paper has broken up the face image using wavelet packet and then usedPCA to make the vector smaller than before. So the paper mainly o?ers two aspectsof works as follows:(1)The original LDB selects the best bases and its coe?cient features with thecriterion of the di?erence of class energy, but the criterion is not the best one inpattern classification. This paper studies several separability criteria.(2)The dimension of the vector which gotten from LDB is also very large.Andsome coe?cient in the vector is useless or harmful to recognition.So the paper hasused PCA to solve the problem.The method has improved the recognition resultand made recognition speed faster than before.The paper has used the ORL database.Broken up the database to two parts andeach of them has two hundred images of twenty people. The method has gotten theaverage recognition rate: 91.18% for EigenFace,91.84% for original LDB,92.45%for the method of the paper.And the method made the recognition speed faster thanbefore. |
PDF Full Text Request