摘要
针对Adaboost人脸检测算法在分类器训练过程中耗时较多的问题,对Adaboost算法进行了详细分析,提出了加快寻找每一轮最佳弱分类器的四点均值法。该方法对每个特征,计算所有训练样本对应的特征值,并将其从小到大排序,求相邻的4个特征值的平均值,该平均值作为阈值,计算错误率,找出最佳弱分类器。减少特征量,修改弱分类器权重,加快收敛速度,使用不同遮挡部位的人脸样本训练分类器,实现了局部遮挡人脸的检测。实验结果表明,该方法明显提高了训练速度,缩短训练时间,并能较准确地检测局部遮挡人脸。
Aimed at the time-consuming problem of Adaboost face detection algorithm in the training classifier process,a detailed analysis of Adaboost algorithm is carried out,the four-point average method is proposed to speed up and look for the best weak classifier.Using this method,for each feature,the corresponding feature value of all training samples are calculated and ordered from small to large,a average values of four adjacent feature are found,the average is looked as a threshold to calculate the error rate and find the best weak classifier,and reduces features,and modifies the weight of weak classifier to increase the convergence speed.Using different partial occlusion face samples train classifier,partial obscured face detection is realizded.The experimental results show that the method can significantly improve training speed,shorten training time,and accurately detect partially obscured faces.
出处
《计算机工程与设计》
CSCD
北大核心
2011年第3期984-987,共4页
Computer Engineering and Design
