Face Detection System Design And Implementation

Face detection as a special subject of object detection,has been intensively researched in recent years due to their importance in applications such as natural human computer interfaces,content-based retrieval, object based video compression,digital video process and video surveillance.This thesis focuses on how to detect faces in gray or color images at high accuracy,also the way of combining many face characteristic information such as skin color is an effective menthod when requiring the high detection speed. Based on the Gentle AdaBoost——a variant form of AdaBoost which is populared in the area of face detection,a rapid face detection system is designed and realized. This system is divided into two parts,training(or learning) subsystem and detecting subsystem.The aim of training system is to get a structure constructed by many strong classifiers,the performance of the face detection is determined by the form of this structure in some degree.Through some comparision,it shows that many factors influence the form of the structure,such as sample selection,feature selection,kernel arithmetic.This thesis bases on the characteristic of the strong relationship between neighborhood pixels,decreasing the number of rectangle features to reduce the training time.Based on the statistical characteristic of the face samples,picking up the nonface samples from the backup nonface images,this menthod can reduce the number of weak classifiers,quicken the detection speed and maintain the detection performance.The experiment results show that it is successful to use these menthods,the training time can reduce to about twelfth of the original training time,when reaching the fifteenth stage classifiers,the face detection rate can achieve more than ninety precent and the false detection rate largely less than the fixed rate prepared by the system.