摘要
在人脸检测问题中,需要使用风险敏感的AdaBoost算法来最小化人脸的误分类风险。但是现有的风险敏感的AdaBoost算法对位于分类边界附近的低风险样本的分类性能很差,影响了最终的检测性能。为了解决这个问题,该文通过分析风险敏感的AdaBoost算法的分类错误率,从理论上指出了造成该问题的原因,并据此提出了可控风险敏感的AdaBoost算法。经过实验,该算法在相同召回率的情况下比风险敏感的AdaBoost算法取得了更低的虚警率,并且在CMU正面直立人脸测试集上也获得了更优的人脸检测结果。实验结果表明:该算法在保持风险敏感AdaBoost算法优点的同时,提高了对低风险样本的鉴别能力,获得了更好的性能。
The cost-sensitive AdaBoost (CS-AdaBoost) algorithm can be used to minimize the misclassification cost in face detection. However, existing CS-AdaBoost algorithms have difficultly classifying low misclassification cost samples near the classification boundary which influences their performance. This paper theoretically describes how the CS-AdaBoost algorithms' classification error changes near the boundary. A modified controlled-cost sensitive AdaBoost algorithm (CCS-AdaBoost) is then shown to have lower false alarm rates than CS-AdaBoost with the same recall rate, giving a better detection rate in the CMU frontal test set. The results show that this algorithm maintains the advantages of CS-AdaBoost while improving the ability to classify low misclassification cost samples to give better performance.
出处
《清华大学学报(自然科学版)》
EI
CAS
CSCD
北大核心
2012年第12期1703-1708,1714,共7页
Journal of Tsinghua University(Science and Technology)
基金
国家“八六三”高技术项目(2009AA11Z214)
国家自然科学基金资助项目(60972094,61071135)
教育部博士点基金项目(20090002110077)