基于特征数期望最小化的人脸检测分类器构建

目前Boosting训练算法已被广泛地应用于人脸检测中级联分类器的构建,而Boosting及其大量改进算法都主要关注于检测率而不是分类器的性能.文中提出了一种新的基于检测特征数期望值最小化的级联分类器构建方法使得分类器的各层特征数组合达到最佳性能.实验结果表明最优组合的检测特征数期望值比已发表的组合要小将近2倍,从而获得了比已发表的特征数组合高出近80%的性能提升.因此该方法适用于使用Boosting及其变形算法构建具有最佳性能的级联分类器.

Face Detection Based on Minimum Feature Number Expectation

Various training methods using Boosting algorithm to construct a detector cascade are mainly focused on detection rate instead of performance.This paper proposes a novel cascaded classifier constructing method based on the minimization of the cascade feature number expectation which pays direct attention to the performance of detector.The experimental result indicates that the best combination has feature number expectation of 2 times lower than the worst,thus gaining an 80% performance promotion than the published combination.