基于多阈值弱学习的Adaboost检测器

Adaboost detector based on multiple thresholds for weak classifier

近年来基于Adaboost的人脸检测算法因其快速和可接受的检测率得到了成功的应用。但采用单阈值作弱分类器显得太弱难于适应复杂的统计分布,且训练过程较慢收敛。为克服这些困难,采用分类树作弱学习器,该学习器以贪婪的的方法用误差测度减少最大化的划分准则划分节点,并由此生成弱分类器,然后采用RAB或GAB方法在给定数据和标签的训练集上将这些弱分类器提升为强分类器。实践结果表明采用多阈值作弱分类器能显著提高分类器性能。

Recently the human face detection system based on Adaboost is successfully used in application areas because of its high speed and accepted detection rates.However,the Adaboost algorithms using the single threshold weak classifiers are too weak to fit complex distributions,and the training procedure is hard to converge.To overcome this dilemma,this paper provides classification trees as a weak learner.The learner greedily splits the node which causes the biggest reduction in measure of error as the partition criteria and builds a weak classifier.Then boosts a weak classifier using real Adaboost or gentle Adaboost methods on training dataset given in data and labels.Experimental resuhs show that using multiple thresholds for the weak classifier can improve the performance of the classifier significantly.