Research On Discriminant Feature Extraction Of Human Face And Classifier Design

Feature extraction and classifier design are two important parts of pattern recognition system, so are in the field of face recognition. Combining the existing feature extraction theory with transformation methods is a primary research field for effective feature extraction, and applying the artificial neural networks to classification of face features is one of the effective approaches. Feature extraction theory with manifold methods and classifier based on neural networks are further discussed in this dissertation, in which some proposed algorithms work well for human face recognition.Wavelet transform method is widely used in digital image feature extraction and compression for its excellent localization ability in the time and frequency domain. The KPCA(Kernel Principal Component Analysis) can transform the original space into a high-dimension space with some non-linear projecting function, and then extract the principal components in the new space with kernel tricks. Considering the distribution of shape and texture features of face image, a feature extraction method of intersected human face based on wavelet transform and KPCA is proposed. With this method, the human faces are firstly divided into different pieces of small images and then transformed with wavelet transformation algorithm, and different frequency coefficients are chosen as extracted wavelet features according to the position of intersected small images. Then through proceeding to extract the principal components of these features with KPCA and combining them, the new discriminant features are obtained. The experimental results on ORL and Yale face database show that the proposed method is superior to traditional PCA methods, as well as fairly robust to the variety of different illumination condition, face pose and expression.Owing to the strong feature describing ability of image's singular value, the tradition method of singular value decomposition (SVD) is improved, and a feature extraction theory with SVD threshold compression which can retain image's essential information is proposed in this dissertation. Then on the base of extending feature extraction methods of frequency domain and combining them with wavelet transform and discrete cosine transform of intersected human face, the theory based on frequency domain transform and SVD threshold compression is proposed in this dissertation. The experimental results on ORL and NUST603 face database show that the proposed method is superior to some popular face recognition algorithms, such as transient SVD method, and it's also fairly robust to the variety of size scale, different illumination condition, posture and expression of face image.Besides effective dimension compression, the invariable face feature which is immune to face rotation, posture change and illumination condition should be also extracted. On the premise of some research for face invariable geometrical features, a feature extraction theory based on self-adaptation...