The Research On Face Detection Based On General Sample Adaboost

Face detection has gotten great development since 1960s when it was a part of face recognition.After the rapid advancement of electronic commerce,however,human face has become one of the most important recognitional object,then face detection became a independent project.After that time,many algorithms of this area had been brought out,the Adaboost based on cascade invented by Viola and Jones in 2001 is one of the most excellent one, because it is the first real-time face detection method whose detection rate is very high.But asymmetric,the specialty of face detection, has not be figured out very well,What's more,the training speed of this algorithm is very slow.Futurnities make improvement on Adaboost from this two area.This paper betters the accuracy by changine training method and detection process. In this article,the main research work is as follows:The paper focuses on Adaboost,after its theoretical basis,work flow and some improvement based on Adaboost face detection are introduced,a new face detection algorithm is adavanced:1. The paper proposes general sample face detection algorithm based on Adaboost.In the real world,every picture has its own detection difficulty and most of the detection rates are moderate. This paper presented an advanced algorithm—general sample Adaboost by taking use of this law which focuses on the samples whose detection rates are moderate.First, the method defined the scope of general sample by analysising training sample.During training, the weight updating method for the general samples is specially designed,and the non-general samples use the general method.The experiment results show the new method gets better detection rate.2. Better detection framework.In the old framework,all the face must be detected by all the cascade classifier,a rejection by any classifier could lead to a false rejection,the false rejection rate was high.In order to lower the false rejection rate to get better detection rate,the paper designs a second detection function, the method cannot be used unless the false rejection rate is very high.The experimental results show this method can get better detection rate.In the end,a system is designed to realize the training and the detection process.