AdaBoost分类器的一种快速训练方法

A fast training method for AdaBoost classifier

针对训练AdaBoost分类器的计算量随候选特征和训练样本数量的增加而急剧增加问题,提出了AdaBoost分类器的快速训练方法.AdaBoost分类器由多个决策桩构成.由于正负样本特征值分布的随机性,现有方法都在训练样本的特征值中穷举搜索来获得最佳决策桩.首先,注意到优秀特征阈值-误差(T-E)曲线的近似凸性,提出使用二分搜索法确定最佳决策桩.与穷举搜索相比,比较操作时间复杂度由O(N)降低为Olog N AdaBoost分类器的快速训练方法.在公开行人检测数据集Inria Pedestrian dataset和Caltech Pedestrian Detection Benchmark上的实验表明,提出的快速训练方法得到的分类器与普通方法的检测性能相当.

To resolve the problem that the computation of training AdaBoost classifier increases sharply with the increase of candidate features and training samples,a fast training method for AdaBoost classifier is proposed.An AdaBoost classifier is composed of multiple decision stakes.Because of the randomness of the features values distribution of positive and negative samples,the existing methods exhaustively search the feature values in training samples to obtain the best decision stakes.First,based on the approximation convexity of the excellent feature’s threshold-error(T-E)curve,the binary search method is proposed to determine the best decision stakes.Compared with the exhaustive search method,the time complexity of comparison operation is reduced from O(logN).Secondly,with the revised function of error,the time complexity of the estimation classification error is reduced from O(N)to O(1).Finally,combining the above two,the fast training method of AdaBoost classifier is given.Experiments on tow public pedestrian detection image dataset,Inria pedestrian dataset and Caltech pedestrian detection benchmark show that the classifier trained by the fast training method is peer to that of the ordinary method on performance.