| Face detection is an important research theme in Computer Vision,and it is the foundation of many applications of face image analysis.In real life,human faces have varied poses,so a face detector of a single view cannot meet the demand.This article mainly studies multi-view face detection,starting with the single view detector,and then studies the structure of multi-view face detector,and analyses the performances of different kinds of multi-view face detector through experiments.The main contents of this paper are as follows:Construction of a single view detector.Firstly,this article analyses the advantages and disadvantages of several commonly used facial feature descriptors,and describes the quick calculation method for the extended Haar-like features.Then,the procedure of training classifiers using Gentle Adaboost algorithm is described in detail,and the advantage of Gentle Adaboost algorithm is analysed.After that,it gives the way to train a cascade classifier to accomplish the detection task of single view faces.Finally,some experiments are done to verify the performance and shortcomings of single view detector,and to show the necessity of multi-view detection.Design of multi-view face detector.Firstly,the view space of out-of-plane rotation of[-90°,90°]is divided into 5 categories according to the facial topology distance.Then,the characteristics of 4 kinds of common structures of multi-view face detectors are analysed,and 4 structures of multi-view face detectors are designed.Contrast experiment of multi-view face detector.Several experiments are done on the same test set using the 4 detectors designed before,and the different performance of them on many criteria such as detection speed,detection rate,false alarm,are compared.The reasonability of view division is also verified.Finally,the conclusion of using what kind of detector under certain circumstances is given. |
