基于卷积神经网络的遮挡人脸检测  

Occluded Face Detection Based on Convolutional Neural Network

作　　者：郭灼豪 张镇宇 桂庆丰 张成 丁凡[1] GUO Zhuohao;ZHANG Zhenyu;GUI Qingfeng;ZHANG Cheng;DING Fan(School of Intelligent Engineering,Shaoguan University,Shaoguan 512005,Guangdong,China)

机构地区：[1]韶关学院智能工程学院,广东韶关512005

出　　处：《韶关学院学报》2023年第6期13-17,共5页Journal of Shaoguan University

基　　金：韶关市科技计划项目“智能监控系统中基于深度学习的行人重识别技术应用研究”(200721124530850);韶关学院校级科研项目“基于深度学习的遮挡人脸检测”(SY2019ZK02)。

摘　　要：针对传统人脸检测方法难以检测遮挡人脸,而现今遮挡人脸检测方法又忽略正常无遮挡人脸的检测问题,构建一个新的卷积神经网络AVGNet和一个包含33 w张图像的多元数据集DYFace,用于兼容无遮挡人脸检测的遮挡人脸检测识别.AVGNet改进了AlexNet网络大卷积核和VGGNet网络多参数的缺点;多元数据集DYFace包括口罩遮挡人脸、墨镜遮挡人脸和正常无遮挡人脸.实验结果表明,使用DYFace训练AVGNet,然后运用AVGNet检测人脸,不仅能检测受口罩和墨镜遮挡的人脸,而且不丧失正常无遮挡人脸的检测能力,能在实验集上达到78.0%的正确检测率,分别比传统的AdaBoost和Seetaface高52.0%和32.0%.Aiming at the addressing problem that traditional face detection methods have difficulty detecting occluded faces and current occluded face detection methods ignore unoccluded face detection,a new convolutional neural network,AVGNet,and a multivariate data,DYFace,containing 330k images are constructed for occluded face detection compatible with unoccluded face detection.The AVGNet overcomes the shortcomings of large convolution kernels of the AlexNet and excessive parameters of the VGGNet.DYFace includes images with faces occluded by masks,by sunglasses and with no occlusion.AVGNet is trained with DYFace,then is used for occluded face detection.AVGNet is trained with DYFace,then is used for occluded face detection.The experimental results show that the method can not only detect faces with no occlusion,but also faces occluded by masks and sunglasses at a detection rate of 78.0%in experimental set,which is 52.0%and 32.0%higher than those of AdaBoost and Seetaface,respectively.

关 键 词：卷积神经网络 遮挡人脸检测 多元人脸数据集 

分 类 号：TN29[电子电信—物理电子学] TP183[自动化与计算机技术—控制理论与控制工程]