Research On Tensor Subspace Face Recognition Algorithm

Face recognition is a task which is intensive studied in application of pattern recognition. An intact pattern recognition system consists of feature extraction and classification. The feature extraction is mainly studied in this dissertation. The subspace algorithms are among the common feature extraction algorithms. Especial in recently, the tensor subspace algorithms have been studied more extensively than before. A new trend of feature extraction algorithm is tensor subspace combined with manifold learning. The merits of tensor subspace algorithms and other algorithms are combined in my paper, so some new face recognition algorithms are put forward in my paper. Another new face recognition based tensor subspace and manifold learning is put forward based on overcoming some defects of the current tensor subspace. The face data are consider as vectors in the traditional subspace algorithms, which has its essential defects in face recognition. The essential defects have been reduced in traditional tensor subspace algorithms. But the defects in traditional tensors subspace algorithms are in existence. So the idea of spatially smooth subspace manifold learning algorithm is introduced in this dissertation. And some new face recognition algorithms based on spatially smooth subspace manifold learning are put forward in this dissertation.In sum, the main research results achieved in this dissertation are given as follows:1. The feature extraction effect of principal component analysis (PCA) is better than the effect of principal component analysis. And wavelet has two abilities to capture localized time-frequency information and to reduce the dimension of images. According to the two advantages of the above algorithms, a new face recognition algorithm based on wavelet transform and tensor PCA is proposed. Wavelet transform is firstly used and then tensor PCA is used to extract the feature of sub-band images, and the efficient recognition of face images can be realized.2. The PSO and wavelet combined with tensor PCA algorithm. The extracted features which have been extracted with wavelet combined with tensor PCA algorithm are further optimized. The procedure of the algorithm is features of each face image are firstly extracted with wavelet and tensor principal component analysis algorithm. And every weights of the features'element are determined with PSO according to the clustering right rate of each element, so the object of extracting the key features of the faces can be realized. 3. For to solve the defect of the current algorithms based on tensor subspace manifold learning, a novel tensor subspace learning algorithm is proposed in this dissertation named as: tensor local and global projection. The local nonlinear structure of the data manifold that is the local information of the data can be preserved in the algorithm, at the same time the global information of data is utilized. So the discriminant between classes of data in low dimension subspace can be maximized. And the optimal tensor subspace can be obtained by iteratively computing the generalized eigenvectors and projection.4. A new method to form affinity graph matrix of data based on spectral graph is put forward, using the method, two algorithms for face recognition under the restraining of spatially smooth are put forward. The correlation of pixels in images in the two algorithms is considered more sufficiently than in the traditional tensor subspace algorithms, at the same time, the features of the projected subspace based on new affinity graph matrix have the strongest ability of separating data between classes and the weakest ability of separating data within the same class. So the projection subspace features have stronger ability of recognition. The right recognition rates are enhanced by the two proposed algorithms, which is confirmed with experiments.