Face Recognition In Video Images Based On Graph Theory

To extract and recognize human face in video images has been developed rapidly,particularly in video surveillance and identity authentication and other fields.Because of itsproperties of masqueraded,real-time,non-contact and practical, it has become the focus inrecent years.The paper studies and analyzes the preprocessing of the face image,face detection,feature extraction and recognition and so on,which is extracted from the video image. Facedetection is the foundation of recognition,and only in the background it will be extractedfrom the face location and be conducive to the conduct of identification. In the facedetection part, the main work is focused on analyzing and studying detection algorithmbased on skin color detection algorithm. Adaboost-based face detection algorithm can getgood classification results only with a small number of classifiers, and it has a higherdetection rate.But the method is more time-consuming,and it has a higher detection rate andlow false detection rate for the frontal face image and a relatively simple background.Theskin region will be extracted by color segmentation at first,then some non-face area will beremoved through proportional range of face model,and ultimately the face is detectedthrough Adaboost classification.Through experimental analysis, we can conclud that thedetection rate of the proposed algorithm is higher.As a key issue in the face recognition,there are many ways for feature extraction, butsubspace method is an important research direction.The starting point of the dissertation ismanifold learning algorithm of subspace,and the geometry of image space based on graphtheory is the central issue.The artical firstly describes PCA,ICA and NPE,then analysestheir advantages and disadvantages.These three methods can not work well for thediversity information,for example the variance information between models, particular forthe non-linear data,so Neighborhood Structure Preserving Embedding(NSPE) is advanced.NSPE defines two adjacency graphs, one is similarity graph to describe analogousgeometric relationships and the other is diversity graph to describe different geometricrelationships.Differences scatter matrix is given,which can measue different informationand similar information.Finally,we delimit a univocal criteria for feature extraction.TheSimulation experiments confirmed that the algorithm is valid.