非约束环境下的实时人脸检测方法  

Real-time face detection method in an unconstrained environment

作　　者：段燕飞 刘胤田[1] 王瑞祥 咬登国 张航 DUAN Yanfei;LIU Yintian;WANG Ruixiang;YAO Dengguo;ZHANG Hang(College of Software Engineering,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学软件工程学院,四川成都610225

出　　处：《应用科技》2021年第3期21-26,共6页Applied Science and Technology

基　　金：国家自然科学基金项目(61203172);四川省科技计划项目(20ZDYF0008).

摘　　要：在不受限制的场景下,人脸检测大多部署在移动端或者边缘设备上,这些设备计算能力低且内存较小。使用无锚检测思想,基于CenterNet无锚检测器,提出以人脸为点的检测方法。不需要人工预先设置锚框,也不需要使用非极大值抑制来后处理,模型变得简单且高效。模型的骨干网络使用改进的ShuffleNetV2,轻量级网络保证了模型参数少,性能高效,可以部署在边缘设备或者移动设备上。使用人脸检测数据集WIDERFACE来训练网络模型,这个是数据集包含了多种环境下的不同尺度大小的人脸,适用于非约束环境下人脸检测。此方法在FDDB上实现了超高精度(平均精度值(AP)为97.6%),实现了速度和精度的平衡。In unrestricted scenarios,face detection is mostly deployed on mobile terminals or edge devices,which have low computing power and small memory.A face point detection method is proposed,using the idea of anchorless detection and based on CenterNet anchorless detector.The model is simple and high-efficiency without manual presetting of anchor frames or post-processing with non-maximum suppression.The backbone network of the model uses the improved Shufflenetv2.The lightweight network ensures that the model has few parameters,high performance,and can be deployed on edge devices or mobile devices.Face detection data set WIDERFACE is used to train our network model.This data set contains faces of different scales and sizes in a variety of environments,which is suitable for face detection in an unconstrained environment.The method achieves high precision(average precision value(AP):97.6%)on FDDB,and achieves the balance between speed and precision.

关 键 词：人脸检测 无锚检测器 实时检测 神经网络 轻量级网络 非约束环境 关键点检测 多任务 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]