基于级联Adaboost与示例投票的人脸检测

Face Detection Based on Cascade Adaboost and Exemplar Voting

传统的人脸检测算法在复杂背景、极端光照等非控条件下进行人脸检测的误检率较高。为有效降低误检率,文中提出一种级联Adaboost和示例投票的人脸检测方法。首先采用基于LBP特征的Adaboost算法初步定位人脸可能存在的区域,然后通过人脸示例集建立字典,使用稀疏编码的方法利用示例人脸对这些候选区域进行中心位置投票,根据得票数得到判别结果,排除非人脸区域,最终完成人脸检测。该方法的创新在于将基于字典学习的稀疏编码和基于部件模型的目标检测相结合,级联传统的Adaboost算法,实现非控环境下的人脸检测。在两个数据集上的实验结果表明,该方法在保持较高检测率的同时,有效降低了误检率,且鲁棒性较好。

In the conditions of complicated backgrounds and extreme illumination, face detection based on Adaboost algorithm usually has a higher false positive rate. Present a cascade of two algorithms in this paper, Adaboost and exemplar-based voting, to detect face in static images which is able to reduce the false-positives efficiently. This method utilizes LBP as features and a cascade Adaboost classifier is used to detect faces, and a voting method based on sparse coding is used as the final classifier to verify face or non-face. The innovation of the proposed method lies in combining sparse coding and part based model for face detection. The experimental result shows that this method can detect face with high detection rate, suppressing the error detection rate, with high robustness.