The Research Of Face Recognition Based On Algebraic Features

Face recognition, as a kind of biometric recognition, has a widely application in the public security, authentication, automatic tracking, digital video processing and so on. With the development of modern computer technology, face recognition technology has gradually become the focus of the topics in the field of artificial intelligence research.Face recognition algorithm based on algebra features received wide spread attention and application due to its relatively low computational complexity. However by the limited of external conditions, such as the changes of illumination and expression, there is a prodigious room for improvement and enhancement of the recognition of this algorithm.Based on the existing face recognition algorithm based on algebraic features, researching and analysis from two aspects of improve recognition rate and shorten recognition time respectively for this thesis. These not only effectively enhance the algebraic features of identification accuracy, but the recognition efficiency is also improved prodigiously:(1) An improved PC A algorithm for fast face recognition. An improved PC A algorithm is proposed for fast face recognition. At the first, after a candidate set being selected from all facial images, a Gabor filter is applied to the images in candidate set, meanwhile, the algorithm is used for noise elimination. Then PCA algorithm is implemented for dimensionality reduction and feature extraction on these images. Finally the matched facial image is obtained by using the nearest neighbor classifier. The method used the candidate set of face images which choosing according to certain standards instead of all of the training samples in the stage of preprocessing, feature extraction and dimensionality reduction. This saves a lot of time for processing the original images. Experiments in the ORL database show that the algorithm can not only shorten the recognition time, but also enhance the recognition rate greatly.(2) Fuzzy diagonal2DLDA Algorithm for face recognition. According to the within-class scatter matrix is singular, which makes LDA can not be applied directly to small sample size (SSS) problem, and the existing weakness for locate the center of total samples space inaccurately, we improved the approach to calculate the between-class scatter matrix. Blending the diagonal technology and fuzzy set theorem into2DLDA. Diagonalizing transformation matrices are obtained from input images firstly. Then we calculate the membership degree matrix by fuzzy k-nearest neighbor (FKNN), and incorporate the membership degree into the definition of the within-class scatter matrix and the redefined between-class scatter matrix. Finally, dimensionality reduction and feature extraction have been done by optimal fuzzy projection, and the nearest neighbor classifier has been used to classify.(3) Fuzzy bidirectional weighted sum for face recognition. A new method for feature extraction and recognition, namely the fuzzy bidirectional weighted sum criterion (FBWSC) is proposed in this thesis. FBWSC defines the row directional fuzzy image optimal image projection matrix. Subsequently, each sample in the original training sample set is transformed using the row directional optimal image projection matrix, and the row directional feature training sample set is obtained. Through the fuzzy distance, the row directional weight can be calculated. Similarly, FBWSC defines the column directional fuzzy image optimal image projection matrix; and then obtains the column directional feature training sample set. The column directional weight can be calculated using the fuzzy distance. Having obtained the row and column directional weighted, FBWSC can sum the weighting row and column directional feature training sample sets, and then complete the feature extraction of the original sample data. Experiments on the ORL, FERET and Yale face database show that the proposed FBWSC method for face recognition has high recognition rate.(4) Face recognition using fuzzy discriminant locality preserving projection. FDLPP which is based on maximum margin criterion (MMC), pursues to maximize the difference between the locality preserving between-class scatter and locality preserving within-class scatter instead of the ratio. In FDLPP, fuzzy k-nearest is implemented to obtain correct local distribution information and the pursuit of better classification results. Blending the membership degree into the definition of the Laplacian scatter matrix acquire to fuzzy Laplacian scatter matrix. Experimental results on ORL, FERET and Yale face databases show that the proposed method improved the performance with the change in lighting and viewing directions.