Research On Feature Extraction Methods In Color Face Images

Feature Extraction has been the key problem in computer vision as a hot research, closely connected with face recognition. The performance of the whole system of face recognition is determined by the feature extraction methods directly.The correlation theory about feature extraction methods for face recognition has been made a lot of research in this dissertetion. Main work in this paper can be summarized as following:First of all, the pretreatment system for color face images, include four common color spaces has been studied, the advantages and disadvantages of the face component product feature's in RGB and color space have been stated through many experiments, the component of is better than others for extraction; Three kinds of methods for gray optimization are compared with each other to accomplish the gray-scale processing with color image; in order to highlight the face's detail, the grey level transformation is made; in addition, median fitering is used to eliminate the background noise from the image with noise jamming, on this basis for achieving and improving it using the vector median filtering algorithm for both gray image and color image. Ycb crYcb crThen, the present primary methods of feature extraction for face recognition have been researched: (1) The classic PCA face recognition algorithm is relized and made better, with solving the 2D covariance matrix for image, the more complex calculation for one-dimension could been aviod and the differences about the sample classification has been made full of use, the prccesing of computer the eigenvalues and eigenvectors is simplified, the improved algorithm has better eigenface and dimensionality reduction. (2) Linear discriminant analysis method is also used for testing and improved, the best discriminant features that are most advantageous for classification are extracted by LDA from one-dimension transformation to two-dimensional processing, and then passed the the nearest neighbor classifier classification with Euclidean distance, the correct recognition rate is improved greatly by 95%. (3) The algorithm of face recognition based on kernel principal component analysis can extract nonlinear features of image and can get better performance than PCA especially under less sample training conditions. Two kinds of kernel fuctions: polynomial kernel and radial basis function are choose for experiment by changing the parameter value. Face distribution of low dimension space is mapped to high dimension space by nonlinear transform. Expeiments on different face datasets show that nonlinear face recognition is better than PCA in robustness and actual demand.Lastly, ICA algorithm as the key point is improved by using the maximum entropy for FsatICA, covergence speed is faster, generalized eigenvector is made easily to solve, the use of RKHS dimensional pattern is more flexible, the high recognition rate is reached by 94.5%, which has high practical value.