基于多纹理CS-LBP特征的多视角人脸检测算法  

Multi-view Face Detection Algorithm Based on Multi-texture CS-LBP Features

作　　者：崔凯 才华[1] 陈广秋[1] 谷欣超[2] 孙俊喜 CUI Kai;CAI Hua;CHEN Guangqiu;GU Xinchao;SUN Junxi(Schoog of Egectronic Information Engineering, Changchun University of Science and Technogogy , Changchun 130022, China;School of Computer Science and Technology, Changchun University of Science and Technogog y , Changchun 130022, China;School of Information Science and Technogogy , Northeast Normal University, Changchun 130117, China)

机构地区：[1]长春理工大学电子信息工程学院,长春130022 [2]长春理工大学计算机科学技术学院,长春130022 [3]东北师范大学信息科学与技术学院,长春130117

出　　处：《吉林大学学报（理学版）》2018年第3期610-616,共7页Journal of Jilin University:Science Edition

基　　金：国家自然科学基金(批准号:61172111);吉林省科技发展计划项目(批准号:20160101260JC);吉林省教育厅"十三五"科学技术研究项目(批准号:JJKH20170625KJ)

摘　　要：提出一种多纹理中心对称局部二值模式(CS-LBP)特征,实现复杂环境下的多视角人脸检测.该特征保留Haar特征的序数关系,借鉴局部二值模式(LBP)的组合方式,从水平、垂直、+45°和-45°这4个纹理方向进行特征提取,以保证人脸检测在方向、光照、旋转等方面的鲁棒性.算法采用级联架构,首先针对人脸图像中的不同视角进行分区,分别进行多纹理特征的提取,然后设计独立的分类器,逐级剔除非人脸窗口,最后采用多层感知器(MLP)综合各视角的检测效果,输出检测结果.在数据集FDDB和CMU PIE上进行验证性实验的结果表明,该方法对复杂环境下的多视角人脸检测有效,与传统的卷积神经网络人脸检测方法相比,该方法具有更高的精度.We proposed a multi-texture centrosymmetric local binary pattern（CS-LBP）feature to realize multi-view face detection in complex environments.The feature retained the characteristics of Haar ordinal relations,so we drew on the experience of the combination of local binary pattern（LBP）to extract features from four texture directions,such as horizontal,vertical,＋45°and-45°,so as to ensure the robustness of face detection in direction,illumination,rotation and so on.The algorithm adopted the cascade architecture.First,face images were partitioned according to different angles of view,and multi-texture features were extracted respectively.Then some independent classifiers were designed to eliminate the non-face window step by step.Finally,the multilayer perceptron（MLP）was used to synthesize the detection effect of each angle of view to output the detection results.The results of verification experiments on data sets FDDB and CMU PIE show that this method is effective for multi-view face detection in complex environment.Compared with traditional convolution neural network face detection method,this method has higher accuracy.

关 键 词：人脸检测 积分图 多纹理中心对称局部二值模式 级联结构 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]