Study On Pose-varied Face Recognition Based On Subspace

Face recognition technology is to analyze face images by using computers and extract efficient features which can represent faces. It is one of the important branches of biometric feature identification technology, and has been a hot-point area in the fields of pattern recognition and image processing. Because of its advantages comparing to other biologic features, considerable attention has been paid to face recognition. As this technology can be widely used in many fields in society, it is worthy of further study.Face recognition often meet these problems: the high dimension of samples, the large classes of patterns, a few or a single training samples from each person, pose and illumination variation, etc. And among all the proposed feature extraction methods, subspace methods have been the most popular approaches owing to their appealing properties, such as low time-consuming, well performance on description and separation. Aiming at the condition of facial pose-varied probe images and seldom training samples, this dissertation focuses on the pose-varied face recognition based on subspace methods. The primary contributions of the dissertation are summarized as below:①Aiming at the problem met in applications, that is the recognition rate decreases drastically when the probe samples change, this dissertation proposed the improved face recognition strategy, which is facial pose correction base on sine transform (ST). In the dissertation we added pose correction to traditional face recognition system for transforming the pose-varied face images to frontal face images. This method can keep the texture information of faces, achieve fast facial pose correction, and it is easy to realize. When we only have a few frontal training samples, we can still use those popular methods to guarantee the robustness of system. Comparing with the traditional face recognition system, it will not increase the cost of system after adding pose correction module and at the meanwhile it can increase the recognition rate when encountering pose-varied testing face images.②In practical applications, it is impossible to ask each person to provide many training images. Indeed, in many occasions one person only has one single training sample. However, almost all the face recognition algorithms increase the recognition rate as the training number of each person increase. In this dissertation, some methods have been presented for producing virtual face images from the only given sample. We used polynomial transform and sine transform to generate virtual samples and the recognition rate improved after the numbers of training samples increased. Adding virtual samples can solve the problem of zero class-within matrixes and it makes the subspace algorithms effective on face recognition based on single training sample.③In this dissertation, we combined the subspace feature extraction methods with pose-varied face recognition strategy under some conditions, and verified the recognition performance using some subspace methods after pose correction. The recognition rates have improved.