《厦门大学学报（自然科学版）》

卡口场景下的人脸检测是视频智能监控的关键技术.然而,由于不同的人脸数据集的样本分布之间存在差异,在现有公开数据集上训练得到的人脸检测模型在卡口场景下难以取得令人满意的效果.为了解决上述问题,构建了一个卡口场景下的人脸数据集,并提出了一种简单且有效的模型重训练方法.该重训练方法能在模型检测人脸时,自适应地选取新的训练样本进行模型的重训练.在卡口场景测试集上的实验结果表明,该重训练方法能明显降低聚合通道特征模型的平均漏检率.

Face detection,in videos of passengers going through station ticket barriers,is a fundamental step of the intelligent video surveillance.However,since face data from different datasets follow different distributions,models trained on public-face benchmarks fail to obtain satisfying results in the scene of station ticket barriers.To solve this problem,we first construct our own face dataset in this special scene,and then propose a simple and effective re-training strategy.This strategy self-adaptively selects new samples to retrain a new model when a model is detecting faces.Experiments on test set from the scene of ticket barriers show that this strategy significantly reduces the log-average miss rate of aggregate channel feature model,demonstrating the effectiveness of our re-training approach.

引言
1 构建卡口场景下的人脸数据集
2 ACF模型及其重训练方法
3 实验
4 结论

图1 卡口场景下的示例图像<br/>Fig.1 Example images from the scene of station ticket barriers

图1 卡口场景下的示例图像
Fig.1 Example images from the scene of station ticket barriers

图2 Face++人脸检测器标注训练集(第1行)和手工标注测试集(第2行)中有标注的图像<br/>Fig.2 Annotated images from the training set annotated automatically by Face++(first row)and test set annotated manually(second row)

图2 Face++人脸检测器标注训练集(第1行)和手工标注测试集(第2行)中有标注的图像
Fig.2 Annotated images from the training set annotated automatically by Face++(first row)and test set annotated manually(second row)

图3 ACF模型的原始训练过程<br/>Fig.3 The original training process of ACF model

图3 ACF模型的原始训练过程
Fig.3 The original training process of ACF model

图4 ACF模型的重训练过程<br/>Fig.4 The re-training process of ACF model

图4 ACF模型的重训练过程
Fig.4 The re-training process of ACF model

表1 基于2种负样本采样方式的重训练方法的对数平均漏检率<br/>Tab.1 Log-average miss rate of the re-training methods based on two negatives-sampling approaches%

表1 基于2种负样本采样方式的重训练方法的对数平均漏检率
Tab.1 Log-average miss rate of the re-training methods based on two negatives-sampling approaches%

表2 基于不同的首次训练迭代次数的重训练的对数平均漏检率<br/>Tab.2 Log-average miss rate of re-training based on different iteration times of the initial training process%

表2 基于不同的首次训练迭代次数的重训练的对数平均漏检率
Tab.2 Log-average miss rate of re-training based on different iteration times of the initial training process%

图5 不同模型的实验结果比较<br/>Fig.5 Comparative results of different models

图5 不同模型的实验结果比较
Fig.5 Comparative results of different models

图6 不同模型的检测结果图<br/>Fig.6 Detecting results of different models

图6 不同模型的检测结果图
Fig.6 Detecting results of different models

[1] VIOLA P,JONES M.Robust real-time face detection[J].International Journal of Computer Vision,2004,57(2):137-154.
[2] OJALA T,PIETIKÄINEN M,MÄENPÄÄ T.Multiresolution gray-scale and rotation invariant texture classification with local binary patterns[J].IEEE Transactions on Pattern Analysis & Machine Intelligence,2002,24(7):971-987.
[3] HE N,CAO J,SONG L.Scale space histogram of oriented gradients for human detection[C]∥International Symposium on Information Science and Engineering.Shanghai:IEEE,2008:167-170.
[4] LI J,ZHANG Y.Learning SURF cascade for fast and accurate object detection[C]∥Computer Vision and Pattern Recognition.Portland:IEEE,2013:3468-3475.
[5] Xiao R,Zhu L,Zhang H J.Boosting chain learning for object detection[C]∥IEEE International Conference on Computer Vision.Nice:IEEE,2003:709-715.
[6] WU B,AI H,HUANG C,et al.Fast rotation invariant multi-view face detection based on real adaboost[C]∥IEEE International Conference on Automatic Face and Gesture Recognition.Seoul:IEEE,2004:79-84.
[7] CHEN D,REN S,WEI Y,et al.Joint cascade face detection and alignment [C]∥European Conference on Computer Vision.Zurich:Springer International Publishing,2014:109-122.
[8] BOURDEV L,BRANDT J.Robust object detection via soft cascade[C]∥IEEE International Conference on Computer Vision and Pattern Recognition.San Diego:IEEE,2005,2:236-243.
[9] 严严,陈日伟,王菡子.基于深度学习的人脸分析研究进展[J].厦门大学学报(自然科学版),2017,56(1):13-24.
[10] YANG B,YAN J,LEI Z,et al.Aggregate channel features for multi-view face detection[C]∥IEEE International Joint Conference on Biometrics(IJCB).Clearwater:IEEE,2014:1-8.
[11] 北京旷视科技有限公司.Face++[EB/OL].[2016-11-01].http:∥www.faceplusplus.com/.
[12] DOLLAR P,APPEL R,BELONGIE S,et al.Fast feature pyramids for object detection[J].IEEE Transactions on Pattern Analysis & Machine Intelligence,2014,36(8):1532-1545.
[13] LIU Z,LUO P,WANG X,et al.Deep learning face attributes in the wild[C]∥IEEE International Conference on Computer Vision.Santiago:IEEE,2015:3730-3738.
[14] YU S Q.Libfacedetection:a library for face detection in images[EB/OL].[2016-11-01].https:∥github.com/ShiqiYu/libfacedetection.

备注

引言

1 构建卡口场景下的人脸数据集

2 ACF模型及其重训练方法

3 实验

4 结论

学报简介

备注

引言

1 构建卡口场景下的人脸数据集

2 ACF模型及其重训练方法

3 实 验

4 结 论

学报简介

3 实验

4 结论