Face Recognition Research Based On Non-negative Matrix Factorization

Face recognition is a kind of technology that detect, locate and recognize human faces in images or video streaming. In recent years, the rapid development of computer calculating speed provided better hardware support for image processing techniques, and face recognition technology became a very active research subject in the filed of computer vision and pattern recognition technology. Face recognition technology is extensively used in authentication,visual communication, intelligent monitoring and public security file management.In a broad sense, the process of face recognition includes two parts: the detection and location of human face and the matching and recognition of human face. Firstly, this thesis summarized the research background and research status of face recognition technology, and elaborately introduced the main approaches of human face detection and recognition by means of classification. Then, based on a thorough study of the existing algorithms, this thesis modulated and merged these algorithms together, and designed and completed a face recognition software.Before operating the human face detection, this thesis argued the preprocesing of human face images, which mainly introduced grey level transformation and histogram equalization of human face images. And then,on the basis of the theory, the experiment selected a static image of human face to directly describe the processing effect. In the link of face detection,this thesis adopted Adaboost algorithm to detect human face. Adaboost algorithm resembles a classifier, which regarded human face detection as a two-type classification problem which classifies images into human and non-human groups. It detects human face through cascade classifier after exercising the human face image samples. Then the face detection software based on Adaboost algorithm is completed. This thesis also conducted statistical experiment using the image sample selected from ORL and MIT face database, and observed the influence between the quantity of cascade classifier and the detection rate.In the human face recognition part, this thesis initially introduced the principal of NMF algorithm and some improved algorithms based on NMF algorithm. At last, after analyzing the non-negative matrix factorization and some other improved algorithms, sparsity constraint non-negative matrix factorization algorithm was applied to do the feature extraction, and then discussed the impacts of the changes of the basic matrix sparsity on human face recognition rates, and conducted experiments in ORL face database.The human face recognition and detection experiments in this thesis and the software designed at last are based on Windows system and completed by using Visual C++ 6.0 and OpenCV. The experiment results indicated that the algorithm used in this thesis meet the requirements of the detection rate and recognition rate.