Face Recognition Based On SIFT

Recent years, with the fast development of computer technology, the Face Recognition (FR) has made much progress and wide applications. In widened meaning, the FR actually comprises a series of relevant technologies making up the FR System, including Face Image Acquisition, Face Location, FR Pre-processing, Identity Check and Identity Finder, etc; and yet the FR in narrowed meaning refers in particular to a certain technology or system that can find or check identities by face.The advantage of FR lies in the characteristic that this technology is un-artificial and the detected people would not be aware of it. At present, all kinds of FR devices have been widely used in lots of practical fields, such as the Enterprise Entrance Guard System, Student Time and Attendance System, Police Supervising System and Intelligence Camera, etc. Best of all, in 2008, the FR technology has firstly occurred in Olympic Games and played an important role in safeguarding the opening and closing ceremonies of Beijing Olympics'Games and Paralympics'Games. Although there are so many successes, the FR technology still has a lot of problems, which can be reduced to two classes: inter-class difference and intra-class difference. The first mentioned of two classes is that the face differentiating in different people, the other is that the face variations from one person, such as the variations resulted from different conditions of face expressions, visual angles, lightening, shading and aging. As the result, the SIFT algorithm, proposed by Pro. David Lowe in recent years has gained the attention of FR researchers for its capability of extracting high distinctive features.After the comprehensive analysis on the current situation of FR research home and abroad, this paper, based on the wide application of the local feature operators in FR field, seeing that SIFT operator can work very well on the recognition and matching of rigid objects, has made some hard research trying to apply the SIFT operator with the distinctive competencies to FR system and solve the hot problems existed at the present stage, such as: distortion, occlusion and illumination, etc.The achievements of this paper have presented in two aspects as below:1. To present an improved algorithm using SIFT descriptor combined with a pattern matching strategy based on K-means clustering for FR, furthermore, to adopt an optimized similarity computing method which is combined the vector space similarity with the keypoints matching similarity to calculate the global similarity matching for face images, and make use of the square of similarity and the weight-assignment solution based on probability statistics to improve accuracy in local feature matching. The experiments'results have proved the robustness and effectiveness of our method.2. To introduce a new descriptor which extents SIFT----GLOH into the FR algorithm, that is to change the location grid and use PCA to reduce the size so as to increase the robustness and distinctiveness of it. As showed in experiments of the front face images, the GLOH works best when dealing with all kinds of interfaces, and can get the better results than SIFT algorithms from the distortion and occlusion.