遮挡条件下的人脸检测与遮挡物属性判识  

Face Detection and Obstacle Attribute Identification under Occlusion

作　　者：魏丽[1] 王洁[2] 姜昕言 丁萌[2] WEI Li;WANG Jie;JIANG Xin-yan;DING Meng(Jincheng College,Nanjing University of Aeronautics and Astronautics,NanjingJiangsu 210156,China;College of Civil Aviation,Nanjing University of Aeronautics and Astronautics,Nanjing Jiangsu 210016,China)

机构地区：[1]南京航空航天大学金城学院,江苏南京211156 [2]南京航空航天大学民航学院,江苏南京210016

出　　处：《计算机仿真》2020年第9期441-445,450,共6页Computer Simulation

基　　金：江苏省高校自然科学基金(18KJB590002);国家自然科学基金(U1633105);南航金城学院院级科研项目(2017-y-04)。

摘　　要：针对智能视频监控中具有饰物遮挡的人脸检测问题,采用将级联卷积神经网络与稀疏分类及颜色识别相结合的方法。首先基于级联卷积神经网络的多任务联合模型检测出人脸及双眼、鼻尖和嘴角位置。其次在KSVD字典学习基础上,结合分类标签信息以提高字典的判别性,同时通过一个线性分类器对性别以及是否戴有帽子、眼镜、口罩进行分类,最后根据已经确定的眼睛及嘴的位置大体确定帽子、镜片和口罩区域,并基于HSV颜色空间对饰物颜色进行识别。经过实验室数据库和网络资源图片的大量实验证明,上述方法在不同拍摄角度、不同遮挡情况下,均能达到较好的检测与判识效果。Aiming at the face detection problem with ornament occlusion in intelligent video surveillance,a meth⁃od combining cascade convolutional neural network with sparse classification and color classification is proposed.Firstly,the multi-task joint model based on cascade convolutional neural network was used to detect the position of the face,eyes,nose and mouth.Secondly,in the framework of the KSVD dictionary learning,the information of clas⁃sification label was used to improve the discriminability of the dictionary,and a linear classifier was used to discrimi⁃nate the different genders and classify the images where there were the hats,glasses,and masks.Finally,according to the positions of the eyes and mouth,the ornaments including hats,lenses,and masks were detected.In addition,the color of these ornaments was recognized by the information from HSV space.Experiments based on laboratory data⁃base and a large number of pictures from the Internet show that the proposed method can achieve good detection and recognition in different shooting angles and occlusion.

关 键 词：遮挡下人脸识别 属性识别 级联卷积神经网络 字典学习 色彩识别 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]