摘要
针对复杂场景下,小尺度、模糊和遮挡人脸检测精度低的问题,提出一种基于增强并行级联卷积神经网络的人脸检测方法。在主网络SSD的多层特征图上,通过融合前后层特征图增强原始特征图的辨识度。将多个增强特征图组合成附加增强网络,与主网络并行设置,加快对小尺度,模糊和遮挡人脸的检测速度。在训练阶段为主网络和附加增强网络设置两种基于锚框尺寸的损失函数,并通过加权求和的方式对这两种损失函数进行融合。在FDDB和WIDER FACE数据集上的实验结果表明,该方法实现了比当前主流人脸检测方法更高的检测精度。
To solve the problem of low precision of small scale,blur and occlusion face in complex scenes,this paper proposes a face detection method based on enhanced parallel cascaded convolutional neural network.On the multi-layer feature map of the main network SSD,the recognition of the original feature map was enhanced by the fusion of the front and back layer feature maps;the multiple enhanced feature maps were combined into an additional enhanced network,which was set in parallel with the main network,speeding up the detection of small scale,fuzzy and occluded faces;two loss functions based on the anchor size were set for the main network and the additional enhanced network in the training phase,and the two loss functions were combined by weighted summation.The experimental results on the FDDB and WIDER FACE datasets show that the proposed method achieves higher detection accuracy than the current mainstream face detection methods.
作者
朱富丽
杨磊
姬波
Zhu Fuli;Yang Lei;Ji Bo(Henan University of Animal Husbandry and Economics,Zhengzhou 450044,Henan,China;Henan Information and Statistics Vocational College,Zhengzhou 450008,Henan,China;School of Information Engineering,Zhengzhou University,Zhengzhou 450001,Henan,China)
出处
《计算机应用与软件》
北大核心
2020年第11期101-105,111,共6页
Computer Applications and Software
基金
国家自然科学青年基金项目(61502434)
河南省档案科技项目计划(2019-X-06)
河南省教育厅课题(JYB2019320,ZJC19015)。
关键词
人脸检测
卷积神经网络
增强特征图
附加增强网络
多尺度
Face detection
Convolutional neural network
Enhanced feature map
Additional enhanced network
Multiscale
