Currently data of computing technology applied fields are high dimensional, such as computer vision and image processing, text analyzing in information retrieving, data mining and biometric feature recognition. It is of great value to explore potential meaningful low dimensional data structure from high dimensional observation data and then acquire its compact expression. Furthermore, high dimensional data are prone to feature disaster. Subspace classification powerfully processes low dimensional data transformed from high dimensional data by dimensionality reduction technology. Subspace classification is effective in understanding of high dimensional semantics or context, becoming one of the most important topics in pattern recognition and machine learning because of its high value in use and theoretical significance. This dissertation does a lot of research work on subspace classification technology, presents two new algorithms for both linear classification and non-linear classification, and implements an algorithm system that integrate some common methods for face recognition. The classification schemes in this dissertation rely on the technology of pattern recognition, image processing and information processing. The system depending on the powerful computing capability of computers has a good efficiency and practical value. The work of this dissertation includes the following aspects: (1) Doing lots of research work on current technology of subspace classification, introducing the dimensionality reduction methods of both linear and non-linear methods, and analyzing the advantages and faults of these methods. (2) In order to overcome the unrobustness and irreversible shortcomings of tensor projection, the scheme presents a reversible and stable orthogonal tensor projection method in subspace learning, which has a good performance in face recognition and can be used in practice. (3) Proposing a synthesized discriminant projection frame, which integrates both local and global information of each class, thus achieves better results. (4) Relative system designing of face recognition are introduced.