Abstract: This paper introduces a method of face detection based on the advanced Adaboost algorithm. The method is composed of two phases： training and detecting. The training phase contains three parts： gain the rectangular features based on Haar_Like function, generate the strong classifier by using the advanced Adaboost algorithm, generate the strong face classifier by cascading the strong classifiers. In the detecting phase, the paper introduces a method using the Pyramid Exhaustive Search（PES） method to do face detection on the image which needs to be detected. In order to address the degradation problems which may arise during the training process in traditional Adaboost algorithm, this paper defines a threshold HWt in each round of training. The weights are updated by the situation in which sample is wrongly classified or not or the current weight is greater than the right HWt or not. This approach can avoid the degradation of a serious distorting weight distribution which may appear in the training phase and improve the detection efficiency. After programming practice, the result shows that the method achieves high detection efficiency and better detection accuracy.

Key words: human face detection, advanced Adaboost algorithm, weight distribution, rectangular features, Pyramid Exhaustive Search（PES）, integral image, classifier

摘要： 介绍了一种建立在改进型Adaboost算法基础上的人脸检测方法，整个方法分为训练和检测两个阶段。训练阶段包含提取类Haar_Like矩形特征、利用改进型Adaboost算法生成强分类器、级联强分类器生成人脸检测器三步。检测阶段，采用金字塔式的穷举搜索法将对待检测图像进行人脸检测。为了解决传统Adaboost算法在训练过程中可能出现退化现象的问题，在Adaboost每轮训练中，定义一个阈值HWt，结合样本是否被错误分类以及当前权值是否大于HWt来给样本更新权值，该方法可以避免训练中可能出现的权重分布严重扭曲的退化现象，提高检测效率。经过编程实践，结果证明该方法检测效率高、检测精度较好。

关键词: 人脸检测, 改进型Adaboost算法, 权重分布, 矩形特征, 金字塔式穷举搜索法, 积分图, 分类器