摘要
针对于自然场景下人脸检测存在的姿态复杂、遮挡和光照等问题,提出一种基于4级级联全卷积神经网络的人脸检测算法。构建4级级联网络,采用级联分级训练代替端到端训练,以避免只共享1个网络权值的局限,进而获得有区分性功能的深度网络,提高检测精度;每级深度网络结构均采用全卷积结构,可以接受任意尺寸图像的输入,提高检测效率;另外在训练过程采用自举法Bootstrap进行网络模型的优化训练,提高训练样本利用率;利用最终训练好的深度卷积网络模型实现人脸检测。人脸检测实验结果标明,本算法在自然场景下,对多姿态、遮挡、单图多种人脸类型等均具有良好的鲁棒性,同时在现有平台上每张图片的检测速度达到96ms,在国际权威的人脸检测公开测试集FDDB上的"真正率"达到82.98%。
Aiming at the problem of various facial gestures,occlusions and illumination in human face detection in natural scene,a face detection algorithm was proposed based on full convolution neural network of four-level cascading.Firstly,a four-level cascade network was constructed.A cascade training was used instead of end-to-end training to avoid the limitation of sharing only one network weight.A deep network with differentiated functions can be obtained,to improve the detection accuracy.Secondly,each level of depth network architecture using full convolution network structure can accept any size of image input to improve the detection performance and efficiency.During the training process,the bootstrap method was used to optimize the training of the network model,to improve the training sample utilization rate.Finally,the trained deep convolution network model was used to achieve face detection.The experimental results show that the face detection algorithm has good robustness in the natural scene,such as multi-gesture,occlusion,single figure multiple faces.On the existing platform,the speed of each picture detection reached 96 ms.The accuracy on the international authoritative face detection data set(FDDB)benchmark is 82.98%.
作者
石学超
周亚同
韩卫雪
SHI Xuechao;ZHOU Yatong;HAN Weixue(School of Electronic and Information Engineering,Hebei University of Technology,Tianjin 300401,China)
出处
《铁道学报》
EI
CAS
CSCD
北大核心
2019年第1期80-86,共7页
Journal of the China Railway Society
基金
中国博士后科学基金(2014M561053)
河北省自然科学基金(F2013202254)
2015年度教育部人文社会科学研究项目(15YJA630108)
关键词
人脸检测
4级级联网络
全卷积网络
自举训练
深度学习
face detection
four-level cascade network
full convolution network
bootstrap training
deep learning