Face Recognition Based On WPMMC In The Transformed Domain

Human face is our primary focus of attention in social intercourse, playing a major role in conveying identity and emotion. Face recognition is one of most active research subject in the area of pattern recognition and computer vision. Feature extraction techniques are widely employed in face recognition system to reduce the dimensionality of data and to enhance the discriminatory information. Principal component analysis (PCA), Independent component analysis (ICA), Linear discriminant analysis (LDA) and Maximum margin criterion (MMC) have been widely employed in face recognition system in the past few decades. PCA and ICA neglect the between class distribution and within class distribution, so they are not optimal in searching for the most discriminant features. LDA is based on Fisher criterion and has been proved to be much more effective than PCA in many applications. But it is not stable due to the small sample size (SSS) problem. MMC is proposed in the recently years to avoid the SSS problem of LDA.Discrete cosine transform (DCT) and Discrete wavelet transform (DWT) have been employed in face recognition for dimensionality reduction. Although they are not as efficient as PCA in dimensionality reduction, the advantage of DCT and DWT are that the basis images are only dependent on one image instead of on the entire set of training images. Systems based on DCT and DWT should not be retrained when new classes are added to obtain optimal projection results. In other applications DCT and DWT can be first employed to remove redundant information and the PCA or LDA can be subsequently implemented in the DCT or DWT domain such that the computational complexity can be significantly reduced. It has been proved that the PCA and LDA can be applied in the DCT and block DCT domain and the results are exactly the same as the one obtained from the spatial domain.The main contributions of this thesis are as follows:1) we have learnt the advantages and disadvantages of the traditional statistic feature extraction methods and introduced a generalization of MMC, Weighted pairwise maximum margin criterion (WPMMC), and proved in reality that it is more robust and effective in searching for the most discriminant analysis information. 2) we demonstrate that WPMMC can be directly implemented in the orthogonal transformed domain via theory and application. As DCT, block DCT and DWT transformation matrices are orthogonal, the result of WPMMC in the DCT and DWT domain are the same as the one in the spatial domain. (1) If the databases are stored as compressed JPEG/JPEG2000 images, DCT/DWT coefficients of JPEG/JPEG2000 images can be directly used such that the inverse DCT/DWT can be skipped to cut down computation cost. (2) In JPEG/JPEG2000 images, some redundant information can be removed by quantization such that the dimensionality of image can be initially reduced by removing coefficients with less information.