Research On Face Recognition Approach Using Subspace With Embedding Neighborhood Discrimimant Relation

Biometric technology provides a highly reliable and robust approach to the personal verification. Among all kinds of Biometric technologies, face recognition is a biometric technology with great developable potential, it is believed having a great deal of potential applications in information security, public security and financial security. Face recognition is also one of the most active research area, which is closely related to many disciplines such as Computer Vision, Pattern Recognition, Image Processing,Machine Learning, Cognitive Psychology etc, its research achievements would greatly contribute to the development of these disciplines. The study of face recognition technology has both tremendous application values and important theory significance.During past two decades face recognition technology has made great progress, but many questions still remain unsolved in practical applications. Generally speaking, face image data sampled from observation space are high-dimension and have very complex structures, these data not only exist too high-dimension problem, but also can not clearly reveal object intrinsic nature, therefore data in observation space should not be used directly for classification. To solve the above problem, a suitable feature extraction is necessary for data set. More specifically, by scientifically extracting nature feature data, important underlying structure property in original structure of data set is preserved along with dimensionality reduction. Transferring face recognition classification work from original observation space to feature subspace and reasonable dimensionality reduction can remove noise and speed up the computation, then the recognition process is more effective. The subspace method has been the most popular approach for feature extraction and face recognition.It depends on reasonable selection and construction of features whether the subspace method can efficiently achieve anticipant effect. In practical case, the diverse variations of ultra high-dimension object's observation data were only caused by a limited numerber of environment factors. In analyzing their variation relationship, it is believed that the above high-dimension data do not disorderly scatter in the whole high-dimension space, but lie on some low dimension manifold. The manifold reflects intrinsic structure of the above high-dimension data and it is consistent with the cognitive manifold in human brain. Based on the above theory and understanding, manifold learning theory is highly attracted attention in recent years, and has been developed greatly. The objective of manifold learning is to recover low-dimension manifold structure from high-dimension data, that is to find low-dimension manifold subspace in high-dimension space, compute the corresponding embedding mapping and realize data dimensionality reduction. It is important to correctly embed intrinsic local relation in data dimensionality reduction mapping of manifold learning.This dissertation mainly deals with face recognition subspace approach with embedding neighborhood mapping on the base of manifold learning method. The class discriminant information is embeded in neighborhood mapping, which improves the recognition rate of subspace algorithm. The main works of the dissertation include:1. Research on scatter difference discriminant neighborhood embedding for face recognitionThe subspace methods such as PCA and LDA are the popular method for face feature representation in the past. The subspace methods obtain great research success under ideal condition. These methods mainly consider global linear characteristics of data set, so they are restricted to a great extend while handling practical complex high-dimension data set. Manifold learning theory tries to discover intrinsic structure of data set to get nature research on the problem. Locality preserving projections(LPP) and neighborhood preserving projections(NPP) are respectively linear approximation to laplacian eigenmap(LE) and locally linear embedding(LLE). They both have locality preserving properties, but they don't consider class discriminant information. In this dissertation, on the base of LE and LLE method, a novel subspace method, called scatter difference discriminant neighborhood embedding(SDDNE), is proposed for face recognition. In our algorithm, the neighbor and class relations of training samples data are used to construct the low-dimension embedding submanifold. After being embedded into a low-dimensional subspace, the samples of the same class maintain their intrinsic neighbor relations, while the samples of the different classes are far from each other. In addition, the algorithm also avoids the small sample problem. Experiment results demonstrate that the proposed method yields better recognition rate.2. A nonlinear face recognition method based on orthogonalized kernel locality Fisher discriminant embedding is proposedA supervised method called locality Fisher discriminant embedding(LFDE) is researched. The discriminant vectors got from LFDE are not orthogonal, which makes the scale between the input space and the space spanned by the discriminant vectors change, for this reason we proposed orthogonal locality Fisher discriminant embedding method. Kernel method as a non-linear dimension reduction method is widely used. The ideal of kernel method is derived from support vector machine, and it differs from traditional method that directly reduces the input space into a lower dimensional space. It projects input space into a very high feature space, and performs original linear methods there. The aim of this conduct is to make non-classified problem input space into a classified problem in high dimensional feature space. Because kernel methods only used the inner product of the input sample, the complexity of the computation is not increased. On the base of OLFDE method, orthogonalized kernel locality Fisher discriminant embedding method(OKLFDE) is proposed by skillfully introducing the kernel mapping. OKLFDE inherits the OLFDE's advantage, and effectively extracts nonlinear feature. Experiment results show that OKLFDE method is not only able to simplify the distribution of the face patterns, but improve the final classification effect.3. A face recognition method based on image matrix discriminant locality preserving projections is presentedLike the traditional method such as LDA, Locality preserving projections is based on 1D vector; for the 2D face image, image matrices must be first transformed into 1D vector row by row or column by column. Since the resulting image vectors of face are high dimensional, LPP usually encounters the small sample size problem. In addition, the matrix-to-vector transform procedure may cause the loss of some useful spatial structure information embedded in the original images. The analysis of image directly based on image matrices is more reasonable. Therefore, image matrix discriminant locality preserving projections (IMDLPP) algorithm is proposed. IMDLPP works directly with image matrices which do not need to be converted into vectors. IMDLPP keeps the spatial position information of pixel in face image, and avoids the singular problem. Meanwhile, IMDLPP preserves within-class local structure according to class label information, and adds between-class scatter constraint into the objective function, so that the features have better discriminant characteristics. Experiment results show the effectiveness of the proposed method.