Extraction Of Face Features Based On Bidirectional Two Dimensional Subspace Analysis

Due to its wide rang of application such as financial security, law enforcement and human-machine interface, face recognition is the hot point in pattern recognition field. The key of face recognition is the extraction of face feature. Researchers have proposed many methods for extracting face feature and among them the subspace analysis has been the most popular and successful one. Based on two dimension subspace analysis, this dissertation gives further research for extracting face feature and explore the protection of face feature.The main highlights and contributions of this dissertation are as follows:1) To overcome the shortcoming of two dimensional feature extraction based on one rule, the bidirectional two dimensional feature extraction methods which combine with multiple rules are proposed. Features extracted by projecting images with2DPCA and2DLDA in row and column direction not only describe well the samples but also take full advantage of the discriminant information of the samples which, thus the same class samples are more clustering and the different class samples are more separate in the protected space; Features extracted by projecting images with2DPCA and2DLPP in row and column direction contain not only the global information from2DPCA but also the local information from2DLPP.2) Two dimensional weighted locality preserving discriminant analysis was proposed by weighting the class-between scatter matrix and embedding the nearest-neighbor graphs which characterize the within-class compactness of the same class samples. The proposed method can not only alleviate the overlap of neighboring classes resulted in Fisher criterion for multiple classes question but also discover the submanifold of images space. Further, the more efficient method named (2D)2WLPDAPCA was proposed by combining2DWLPDA with2DPCA for saving memory space and reducing computation complexity.3) Multiple maximum scatter difference (MMSD) is a variant of LDA in which scatter difference replaces Fisher criterion, thus, the "small size sample" question confronted by LDA is avoided in nature. However, MMSD is based on vector which is computationally expensive. To overcome this weakness,2DMMSD was proposed by adopting image protection and furthermore (2D) MMSDLPP is proposed which project images with2DMMSD and2DLPP in row and column direction.4) The difference of gray value of face image in the same region is small, and the redundant information of LBP will increase by comparing simply the current pixel with a specific value to code. In order to reduce redundancy an improved LBP is proposed in which the codes of LBP are obtained by judging the current pixel gray value is in a range or not, thus, the resulted LBP codes can reduce the redundancy. Experiment results in face detection show the proposed method can extract effective face features. A novel image texture feature called local ternary derivative patter was proposed whose codes are longer than LDP and extracting features are of discriminability. Experiment results show also LTDP can obtain higher recognition rates than that of LDP for face recognition task. The recognition rates drop for face methods based on subspace analysis when illuminate change in face is large, this dissertation combines local derivative pattern (improved LBP and LTDP) with subspace analysis to extract face feature. First neighbor mean LBP and LTDP are used to obtain the texture feature of face images, respectively. Then subspace analysis is used to reduce dimension of texture feature. Experiment results show the proposed methods are more resistant to illuminate changes.5) Combining random projection (RP) and two dimensional analysis, a kind of reversible face feature extraction methods is proposed. Projecting respectively face image by RP and two dimension subspace analysis in row and column direction, the extracted features are discriminant, secure and reversible. Various face templates can be obtained by changing random matrix, the random matrix is usually singular and it is difficult to obtain the original face feature. Experiments show that it is difficult for authentication even if the adversary obtains the face image without the random matrix.