摘要
提出了一种基于椭圆肤色模型与可控风险敏感型Adaboost(CCS-Adaboost)算法的多视角人脸检测方法.在人脸检测的离线训练部分,该方法使用Haar-like特征和CCS-Adaboost算法训练样本.CCS-Adaboost在最小化分类错误率的同时能够最小化样本的误分类风险,从而它能够提高分类准确性.在实时检测部分,首先通过使用YCbCr颜色空间的椭圆模型快速检测出可能的人脸区域,然后通过基于CCS-Adaboost的多视角人脸检测器检测人脸.多视角人脸检测器中级联分类器的前四层构成姿态预估部分,如果样本未通过级联检测器的前四层,那么该样本被确定为一个非人脸样本.实验证明该检测器可以有效和准确地检测多视角人脸.
A multi-view face detection method based on elliptic skin-color model and controlled costsensitive Adaboost(CCS-Adaboost)was proposed.In the off-line training part of face detection,samples were trained by using Haar-like features and CCS-Adaboost algorithm.CCS-Adaboost could minimize the total misclassification cost while minimizing the classification error rate,and could improve the classify accuracy of the samples near the classification boundary.In the on-line detection part,the possible face region could be fast detected by skin color detector in the YCbCrcolor space firstly.Then multi-view face detector based on CCS-Adaboost was used to detect face in skin-color region.The multi-view face detector constructed pose pre-estimation by the front four layers of each cascade-type classifier without training alone.If the sample was abandoned in the previous 4layers of the entire cascade-type detectors,the sample was identified as a non-face sample.The experiments demonstrate that the proposed detector can effectively and accurately detect the multi-view faces.
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2015年第S1期271-275,共5页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
