Research On Independent Component Analysis And Manifold Learning

Feature extraction is one of the most elementary problems in the area of pattern recognition. It is critical factor to extract effectively discriminant features for pattern recognition. The process of face recognition can be divided into three stages:face detection and preprocessing of face image, feature extraction and recognition. Recently, studies have shown that, high-order statistics contains much important information, and moreover, the data of each face image is of a non-linear distribution,many methods of feature extraction adopts manifold learning to explore intrinsic structure of face patterns. Based on the smanifold learning in application to face recognition, some relative research was made in this paper. And the main works are presented in the following:1. Muiti-modal Disturbing Face Recognition AlgorithmIn order to overcome the instability caused by face detection and curses of dimensionality resulted from small sample size, we generated four new images from one two-dimensional image by horizontal and vertical translation, and formed a new image database by adding these new images to original image database. Using local structure information, we designed objective function and obtained dual projection matrix based on two-dimensional images. This method eliminate the small sample size problem to some extent, expand the diversity represented by the face image, and help to overcome the image of instability due to displacement. The algorithm constructs the within-class nearest-neighbor graph and the between-class nearest-neighbor graph according to label infomation. Then the local geometry of the manifold structure and the discriminant structure of manifold are maintained. The proposed method has been tested on many image databases. Compared with the traditional face recognition method, it has a better recognition performance and the experimental results on Yale and ORL face image database show that it is effective and robust.2. Bilinear Discriminant Analysis Method Based on Local Structure InformationA new linear discriminant analysis method is proposed based on image matrices. That’s bilinear discriminant analysis method based on local structure information, which extracts the features from both row and column simultaneously. outliers can be seen as the image distortion caused by the introduction of noise. This method defines sample reliability and detects outliers based on sample neighbor information. Degree of reliability also indicates the degree of a sample being a boundary point, so combining it with the sample adjoining weight and reliability to design the feature extracting object function. Using the charateristics of the within-class scatter matrix and between-class scatter matrix to design an iterative algorithm to compute the projection matrix will reduce the complexity of the algorithm. It has a better recognition performance and the experimental results on Yale and ORL face image database show that it is effective and robust.3. Uncorrelated Linear Discriminant Analysis and Face RecognitionThis paper interprets the principle component analysis algorithm, the forming of Eigenfaces space and the achieving process of the Eigenfaces algorithm. It also interprets the subspace linear discriminant analysis algorithm. Then we use image matrix to construct the image scatter matrix directly. On this foundation we begin the discriminant analysis based on image matrices. That is to say, it need not to convert the image matrix into high-dimensional image vector like the existing image-vector-based linear discriminant methods, so much computational time will be saved if the method is used for feature extraction. A new Fisher optimal criterion has been designed to gather the within-class samples and separate the between-class samples, then using its discriminant vectors for face data projection. The discriminant vectors are Uncorrelated each other. The experimental results on ORL and YALE face image database demonstrate its effectiveness.