文献类型：期刊文献

中文题名：基于单一神经网络的多尺度人脸检测

英文题名：Multi-scale Face Detection Based on Single Neural Network

作者：刘宏哲[1];杨少鹏[1];袁家政[2];王雪峤[3];薛建明[1]

第一作者：刘宏哲

通讯作者：Yang, SP[1]

机构：[1]北京联合大学北京市信息服务工程重点实验室;[2]北京开放大学;[3]北京联合大学计算机技术研究所

第一机构：北京联合大学北京市信息服务工程重点实验室

通讯机构：[1]corresponding author), Beijing Union Univ, Beijing Key Lab Informat Serv Engn, Beijing 100101, Peoples R China.|[11417103]北京联合大学北京市信息服务工程重点实验室;[11417]北京联合大学;

年份：2018

卷号：40

期号：11

起止页码：2598-2605

中文期刊名：电子与信息学报

外文期刊名：Journal of Electronics & Information Technology

收录：CSTPCD;;EI(收录号：20190506455175);Scopus(收录号：2-s2.0-85060826956);WOS:【ESCI(收录号:WOS:000456146900009)】；北大核心:【北大核心2017】；CSCD:【CSCD2017_2018】；

基金：The National Natural Science Foundation of China (61571045), The Supporting Plan for Cultivating High Level Teachers in Colleges and Universities in Beijing (IDHT20170511), The National Science and Technology Support Project (2015BAH55F03), The Foundation of Beijing Union University (Zk10201703), The Foundation of Beijing Municipal Education Commission (KM201811417002)

语种：中文

中文关键词：多尺度人脸检测;上下文信息;特征图融合;卷积神经网络

外文关键词：Multi-scale face detection;Contextual information;Feature map fusion;Convolution neural network

摘要：人脸检测是指检测并定位输入图像中所有的人脸,并返回精确的人脸位置和大小,是目标检测的重要方向。为了解决人脸尺度多样性给人脸检测造成的困难,该文提出一种新的基于单一神经网络的特征图融合多尺度人脸检测算法。该算法在不同大小的卷积层上预测人脸,实现实时多尺度人脸检测,并通过将浅层的特征图融合引入上下文信息提高小尺寸人脸检测精度。在数据集FDDB和WIDERFACE测试结果表明,所提方法达到了先进人脸检测的水平,并且该方法去掉了框推荐过程,因此检测速度更快。在WIDERFACE难、适中、简单3个子数据集上测试结果分别为87.9%, 93.2%, 93.4%MAP,检测速度为35 fps。所提算法与目前效果较好的极小人脸检测方法相比,在保证精度的同时提高了人脸检测速度。
Face detection is finding and locating all faces in the input image, and then returning the position and size of the faces. It is an important direction of target detection. In order to solve the problem which is caused by the diversity of face size, a new single shot multiscale face algorithm is presented based on feature fusion. This method combines predictions from multiple feature maps with different resolutions to handle faces of various sizes, and the fusion of the feature maps in the shallow layers can improve the detection accuracy of the small size face by introducing the contextual information. Experimental results on the FDDB and WIDERFACE datasets coafirm that the proposed method has competitive accuracy. Additionally, the object proposal step is removed, which makes the method fast. The proposed model achieves 87.9%, 93.2% and 93.4% Mean Average Precision （MAP） on the WIDERFACE sub-datasets respectively, at 35 fps. The proposed method outperforms a comparable state-of-the-art HR model, and at the same time improves the speed while ensuring the accuracy.