中文摘要：

针对传统AdaBoost算法的不足,分析了训练过程中出现的退化问题及样本权重扭曲的现象,并提出了解决这一问题的有效方法。该方法对样本权重的更新规则进行了适当的调整,即为每一轮循环设定一个权重更新阈值,根据样本是否被错误分类以及当前权重是否大于该阈值来更新样本权重,从而限制了困难样本权重的过分增大。使用该方法训练级联人脸检测器,试验结果表明,该方法较好地解决了传统AdaBoost算法所出现的退化问题,在保证检测率的同时降低了误检率。

英文摘要：

Focusing on the disadvantages of classical AdaBoost algorithm, this paper mainly analysed the issues of overfitting and distortion of sample weights in training process and come up with a new method to avoid the phenomenon of overfitting. The proposed approach set a weight threshold for each loop, and updated weight of sample according to whether the current weight was greater than the threshold, so that weights of hard samples would not expand too large. A cascade face detector was established using the method. The experimental results show that the new method will not lead to overfitting like classical AdaBoost often does, and it will reduce false alarm rate while holding a high detection rate.