基于AdaBoost的弱分类器选择和整合算法

摘要

Aimed at the problems such as the time-consumption with the training and high mis-detection rate in the process of human face detection with traditional AdaBoost algorithm, an improved AdaBoost algorithm was presented. In this new detection algorithm, two phases of weak classifier selection and conformation were improved on the basis of PSO-based AdaBoost algorithm. In the stage of weak classifier selection, some useless features were rejected before iterative selection of the optimal weak classifier with PSO so as to decrease the search space of the particles. In the stage of weak classifier conformation, a new set of initial parameters of weak classifier, which were related with the ability of positive sample selection, was used to tune these parameters optimally with a kernel function-based nonlinear perceptron algorithm. The experiment result of MIT database indicated that, compared to the PSO-based AdaBoost algorithm, the detection capability was remarkably improved by using the algorithm presented in this paper.%针对传统AdaBoost算法在人脸检测中训练耗时和误检率高的问题,提出一种改进的AdaBoost算法.新算法在基于PSO的AdaBoost算法基础上对弱分类器的选择和整合两个阶段进行改进.弱分类器选择阶段,在使用PSO迭代选择最佳弱分类器之前,剔除部分无用特征,缩小粒子搜索空间；弱分类器整合阶段,在采用基于核函数的非线性感知器算法优化调节弱分类器参数的过程中使用一种新的与正样本分类能力有关的弱分类器初始参数.采用基于MIT数据库的实验结果表明,本文算法比基于PSO的AdaBoost算法在检测性能上有明显提高.