Research On The Methods Of Machine Recognition Of Faces

The machine recognition of faces is one of the most important identification recognition techniques based on biological features. It analyzes face images with computer, and then extracts effective recognition information from them, which is used to recognize the identification of a person, and can be applied to surveillance, management and control. It is a practical technique that spans several disciplines such as digital signal processing, image processing, pattern recognition and computer vision. It can be widely used in commerce, justice and public utility. The research on face recognition technique gives not only important practical value, but also important driving effects for the research on the theory in related fields. The research direction has been the spotlight of the research in pattern recognition field.The research of the face recognition technique mainly focused on how to extract the features describing face image. It needs that these features are tolerant of geometry changes, impression changes and illumination changes; meanwhile the features should contain the information that distinguishes this face pattern from others. In this thesis, we will discuss several relatively new face recognition techniques, and propose improved algorithm respectively.Detailed research work contain:The research on face recognition method based on uncorrelated features discriminant vector sets. The face recognition method based on Foley-Sammon transform to derive discriminant vector sets is based on Fisher linear discriminant criteria. The discriminant vectors derived through the above method are linearly uncorrelated, but the projections of face image on those discriminant vectors are linearly correlated. To improve it, someone proposed a face recognition method based on uncorrelated features discriminant vector sets, and gave better recognition performance. By analyzing the process of deriving uncorrelated features discriminant vector sets gained by traditional method, we proposed a new method, gaining uncorrelated features discriminant vector sets, which doesn't need recursive process. Theory analysis and experiment results showed that the method in this thesis is able to get the same uncorrelated features discriminant vector sets as those gotten in traditional method, but the time cost in the method of this thesis is shorter. Further analysis proved that if all the eigenvalues of the matrix were not equal, the uncorrelated features discriminant vector sets would be the same as those derived by Fisher linear discriminant criteria. Using our new method we can also get Fisher linear discriminant vector sets, but it doesn't need to compute the converse of a matrix.The research on face recognition method based on Fisher discriminant vector sets derived from sub-pattern of image matrix.Usually, the dimension of the data describing face pattern is high. For the method like principal component analysis, the computation quantity is large and the computation time is long. Due to that, someone proposed a new face recognition method based on principal component analysis on face image matrix. For this method, the computation quantity is less and the recognition ratio is higher. We analyzed deeply the realizing principle and found that in fact this method uses principal component analysis on the row vectors of a face image to derive face features. Then we proposed a new face recognition method based on Fisher discriminant vector sets derived from row vectors of image matrix. It computes separately face features describing face image from the row vectors of a face image based on Fisher linear discriminant criteria. The experiment showed that the recognition ratio of our method is higher than those based on image principal component analysis. Furthermore, we computes separately face features from the column vectors of the above feature matrix based on Fisher linear discriminant criteria. It got a very ideal recognition ratio. The above method is a special case of face recognition method based on Fisher discrimina...