Dissertation > Industrial Technology > Automation technology,computer technology > Computing technology,computer technology > Computer applications > Information processing (information processing) > Pattern Recognition and devices > Image recognition device

Face Recognition Researches on Large Scale and Each Person with Few Samples

Author ZhangShengLiang
Tutor YangJingYu
School Nanjing University of Technology and Engineering
Course Pattern Recognition and Intelligent Systems
Keywords face recognition pattern recognition feature extraction principal component analysis (PCA) Feature fusion Fisher linear discriminant analysis (LDA or FLDA) two dimensional projection analysis Virtual samples Multilevel Classification
CLC TP391.41
Type PhD thesis
Year 2005
Downloads 815
Quotes 7
Download Dissertation

Face recognition is a research hot-point in the pattern recognition and image engineering. It belongs to Biometrics. Face recognition normally be regarded as have three processes that are face detection, features extraction and pattern classification. Face recognition often meet these problems, the dimension of samples too high, the classes of pattern too many and each person could only provide a few training samples. This paper tried to deal with these matters.Traditional algebraic feature extraction approaches based on vector patterns, so we call them one-dimensional methods. When patterns are not vectors such as facial images, these methods may meet many high dimensions and calculate difficulties.Two-dimension PCA (2DPCA) and other two-dimensional projection methods can directly extract features by using original image matrixes. But the features extracted by two-dimensional approaches are still matrixes; it could cause the magnitude of features too much and slow down the classification speed. Two new algorithms are used to compress the feature matrixes in this paper. The first is called two dimension double projection methods, these feature matrixes are compressed in vertical direction using two dimension methods again; the second method combined the virtues of two-dimension method and one-dimension method, which use one dimension method to compress the feature matrixes after extracted by 2DPCA. In addition, this paper gives a new 2D-PCA, which use original image matrices to compute between-class covariance matrix and its eigenvectors are derived for images feature extraction.Almost all face recognition algorithms increases the recognition rates as the training number of each person increased. But in practical applications, It is not reasonable asked each person provide many training images, indeed much times there are exactly one person only has one training sample. In this paper, some methods are presented for production virtual images from the given sample. A new method it can generate the simulated images of head after rotated an angle was presented. It generalizes the methods based on Fisher rule to one sample per person. The experiments on FERET and ORL face-databases indicate that after adding virtual images the recognition rates increased efficiently.A novel fusion features extraction method is presented in this paper. We use two-dimension method to extract first group of features, denoted by α. Then, using

Related Dissertations
More Dissertations