Face Recognition System And Face Detection Algorithm

Face detection as an object detection problem of the widest application, is the focus of image engineering and pattern recognition. The problem is addressed to be the first step of face recognition system and surveillance system. When the face recognition technology is applied in finance, judicatory, national security, E-commerce and E-government affair, the face detection is spreading more and more widely. Recently the value of fields such as security access control, video surveillance, search based on content and the next generation of man-machine interface has appeared. Face detection has caught researchers' attention independently.As result of face variance in detail, for example, the difference of appearance, expression and skin color, the disturbance of classes, beard and hair, the effect of illumination, it's challenging to detect rapidly the face from restrictive background. Relying on the analysis of existed face detection techniques, this paper advances a series of algorithms from feature exaction to training and detection, and describes an integrated face recognition system.First of all, an improved face detection training method-Neighbor-Eliminated Boosting (NEB) algorithm is proposed. Adaboost algorithm is the most effective application in face detection. However, it has its own limitations: the classifiers based on cascade structure may unbalance on the FRR( false reject rate) and FAR( false accept rate); and the invalidation of monotonicity assumption may result in the failure of learning. NEB algorithm constructs a group of new feature describers linked by two lists, which brings in correlation of features to simplify training. Experimental results demonstrate that the proposed algorithm accelerates the training speed and obtain the better performance.Secondly, a new feature-LBHF is represented. LBHF( Local Binary Haar Feature) combines haar-like features with LBP( Local Binary Pattern) . So when it's easy to calculate and use as haar-like features, the operator can achieve rotation and gray-scale invariance as LBP. It's experimented that the new feature is effective in face detection.