改进的人脸检测训练方法

An Improved Face Detection Training Method

针对AdaBoost存在的诸如分类器的级联结构会导致系统拒真率与认假率的失衡,单调性前提的不成立容易直接造成训练过程的失败等缺陷,对人脸检测训练方法进行研究,提出了一种改进算法——neighbor-eliminated boosting(NEB)算法.此算法通过构建一种新的基于双表链接结构的特征描述子存储结构,引入特征相关信息,简化了训练过程.实验结果表明,以NEB算法为基础实现的人脸检测系统,在训练速度上具有明显的优越性.

Applied to face detection, although AdaBoost is one of effective algorithms, it has some limitations. Neighbor-eliminated boosting（NEB） algorithm is proposed to remedy these deficiencies, which is like that the cascaded stage classifiers may unbalance on false reject rate and false accept rate, and that the invalidation of monotonicity assumption may conduce to abortive feature learning. NEB constructs a group of new feature describers linked by two lists, which will lead to correlation of features to simplify training. Experiments demonstrate that NEB algorithm accelerates the training speed and obtain the better performance.