Research On Face Recognition System Based On Adaboost And Support Vector Machine

Human face recognition is one of the most active and challenging tasks for computer vision and pattern recognition. It can be widely applied in such fields as personal identification, human-computer interface, visual communication, criminal archive administration, content-based image retrieval, etc. A typical automatic face recognition system consists of the following three functions: face detection and location, feature extraction, recognition or verification. This dissertation mainly studies the approaches to real time face detection and recognition in static images. Much improvement has been made in the algorithm's practicality. And then an automated face recognition system prototype is established.Two face detection methods are proposed in our prototype system. One is using Adaboost learning method and the other is based on skin color model in YCbCr chrominance space. The common characteristic of the algorithms is the rapid detection speed and at the same the two methods are complementary.The adaboost method is based on Haar-Like feature Classifier Cascade and AdaBoost learning algorithms recently proposed by Viola and Jones. We choose the extended Haar-like feature after comparison and analyzed how the parameters such as number and arrangement of the Stage Classifiers affected the system' s capacity. A rapid face detection system is designed and realized on the basis of the extension. The system can detect frontal images including a wide range of formats. The detection experiment with the CMU databases shows that the system reached a high hit-rate and low false-alarm-rate. Especially, the detecting speed of the system is very high and almost attains thecriterion of real-time.In skin color model face detection, after lighting compensation a nonlinear color transformation technique is applied to the face images in YCbCr chrominance space. Then we can get the practical skin tone cluster model to detect the potential face. Experiment results show that multi-pose face in color images can be detected.A novel approach to face recognition has been proposed in this dissertation, DCT/SVM combined method. Discrete cosine transform (DCT) is processed to raw face images to get the feature vectors and proper feature selection approach is applied according to the characteristic of the DCT coefficients. A multi-step approach based on the one against one classifying strategy is chosen to extend support vector machine (SVM) capability to deal with multi-class face recognition. Compared with other methods of the experiments on ORL face database, the recognition rate of the proposed method improves a lot.