Multi-View Face Detection And Demographic Classification

In this thesis, a Look-Up-Table (LUT) type weak classifier based realAdaBoost algorithm is proposed and applied to the face detection problem andthe demographic classification problem. This method is an improvement of the threshold type weak classifier baseddiscrete AdaBoost algorithm of P. Viola and M. Jones [8]. LUT weak classifiersare able to fit more complex sample distributions than threshold weak classifiers, andthis makes them significantly superior to threshold weak classifiers since thesample distribution will become more and more intricate during the iterationprocedure in boosting. Real AdaBoost is superior to discrete AdaBoost in learningcapacity since it deals with not only the Boolean prediction but also the real-valueconfidence which represents the classification margin to some extent. Theexperimental results show that the combination of LUT weak classifier and realAdaBoost has achieved a faster convergence speed and a better performance. On the face detection problem, the cascade-structured detectors of P. Violaand M. Jones [8] achieved good performance and real-time speed, while thedetectors learned by our method perform even better. To detect faces with variousposes, view-based face detectors are trained and a novel pose estimation method isdesigned. The detectors achieve much higher correctness on the 360-degreerotation-in-plane face detection and the multi-view face detection problems thanexisting methods as far as we know. Finally an omni-directional face detectionsystem is implemented. The bi-class and multi-class real AdaBoost algorithms are used to learn thegender, age and ethnicity classifiers. Compared with the Support VectorMachine algorithm, the demographic classifiers learned by AdaBoost haveachieved comparable correctness performance with a much higher processingspeed that is up to two grades. In order to implement a demographicclassification system, a facial feature extraction method based on thecombination of SDAM and SFBG (Simple Direct Appearance Model and Simple英 文 摘 要Face Bunch Graph) is proposed that is used for face alignment. As an interestingapplication of demographic classification, a face picture retrieval systemprototype is implemented based on automatic demographic classification. The successful applications on face detection and demographicclassification show that the LUT weak classifier based real AdaBoostalgorithm is a very efficient image pattern recognition method, especially forface pattern classification related problems.