计算机系统应用  2020, Vol. 29 Issue (10): 89-96 PDF

Face Recognition System Based on Front and Back Interaction
HOU Jing-Yan, SONG Huan-Sheng, LIANG Hao-Xiang, JIA Jin-Ming, DAI Zhe
School of Information Engineering, Chang’an University, Xi’an 710064, China
Foundation item: National Natural Science Foundation of China (61572083); Joint Fund of Ministry of Education (6141A02022610); Major Project of Key Research and Development Program of Shaanxi Province (2018ZDXM-GY-047); Team Incubation Project of the Central Universities of China (300102248402)
Abstract: Aiming at the problems of low calculation efficiency and poor robustness of the existing face recognition system, this study proposed a face recognition system based on front and back interaction, including client, database, and server. First, a GrabCut-based facial Region Of Interest (ROI) extraction algorithm was proposed for the client end. Second, the extracted ROI is transmitted to the server, and the ResNet network is used on the server to extract facial feature points according to the ROI. Finally, the facial feature points extracted from the server were returned to the client, and the client performs Euclidean distance matching between this information and the feature points that were pre-stored in the database to obtain the face recognition result. The experiments were performed on the CeleA database and random videos, and the results show that the proposed ROI extraction algorithm significantly improves the accuracy and robustness of face recognition. Moreover, compared with the traditional non-interactive structure, the front and back interactive structure of the system greatly improves the computational efficiency of face recognition.
Key words: face recognition     front and back interaction     residual neural network     feature extraction     GrabCut

ConvNet和DeepFace虽然提高了精度, 但是仍存在一定缺陷, 首先均需要人为选择特征向量, 这将会导致算法具有不确定性, 并且在大型数据库上训练神经网络模型时, 中间层神经单元个数会为了保持特征的完整性而不断增加, 导致最后提取的特征向量维度变高, 增加了计算成本. 其次使用神经网络进行提取全局特征增加了计算机及网络的荷载, 降低了人脸识别的速度.

1)使用了残差神经网络模型(ResNet)[1], 不用手工选取度量方法, 而是将高维度的人脸图像映射为128维特征向量, 再使用欧式距离进行人脸的相似性度量.

2)采用了前后端交互式系统, 提取特征值和特征值距离匹配分别运行于服务端和客户端, 同时通信间采用码流传输格式, 加快系统传输速率.

3)提出了基于GrabCut的人脸兴趣区域(Region Of Interest, ROI)提取算法, 先从系统输入帧图片中提取出人脸前景, 排除了带有冗余和干扰信息的背景, 提升了系统的计算效率和人脸识别精度.

1 基于前后端交互的人脸识别系统结构

 图 1 基于前后端交互的人脸识别系统框架图

2 算法介绍 2.1 基于GrabCut的ROI提取算法

 图 2 图像预处理示意图

 $D(x) = \sum\limits_{i = 1}^k {{\pi _i}{g_i}(x;{\mu _i},{\Sigma _i})} ,\sum\limits_{i = 1}^k {{\pi _i} = 1{\text{且}}0 \le {\pi _i}} \le 1$ (1)
 图 3 图像预处理算法流程图

 $g(x;\mu ,\Sigma ) = \frac{1}{{\sqrt {{{(2\pi )}^d}\left| \Sigma \right|} }}\exp \left[ - \frac{1}{2}{(x - \mu )^T}{\Sigma ^{ - 1}}(x - \mu )\right]$ (2)
 $V(\underline \alpha ,{\textit{z}}) = \gamma \sum\limits_{(m,n) \in C} {\left[{\alpha _n} \ne {\alpha _m}\right]} \exp - \beta {\left\| {{{\textit{z}}_m} - {{\textit{z}}_n}} \right\|^2}$ (3)
2.2 通信过程中的压缩与解压

 图 4 图像压缩与解压

2.3 基于ResNet的人脸识别算法

 图 5 人脸识别算法流程

2.3.1 深度残差网络

2015年由Kaiming He等4名华人共同提出的深度残差网络ResNet (Residual neural Network)[15], 通过增加shortcut connection (Identity Map)来直接连接深浅层网络, 使得梯度能够很好地传递到浅层, 其结构可以加速神经网络的训练, 也大大提高了模型的准确度.

 图 6 残差学习单元

2.3.2 人脸特征提取

 图 7 人脸68个特征点示意

2.3.3 人脸特征匹配

 $dist(A,B) = \sqrt {\sum\limits_{i = 1}^2 {{{({x_i} - {y_i})}^2}} } {\rm{ = }}\sqrt {{{({x_1} - {x_2})}^2} + {{({y_1} - {y_2})}^2}}$ (4)
 $dist(A',B') = \sqrt {{{({x_1} - {x_2})}^2} + {{({y_1} - {y_2})}^2} + {{({{\textit{z}}_1} - {{\textit{z}}_2})}^2}}$ (5)
 $\begin{split} & dist(A'',B'') = \sqrt {\sum\limits_{i = 1}^{128} {{{({x_{1i}} - {x_{2i}})}^2}} } \\ & =\sqrt {{{({x_{11}} - {x_{21}})}^2} + {{({x_{12}} - {x_{22}})}^2} + \cdots + {{({x_{1(128)}} - {x_{2(128)}})}^2}} \end{split}$ (6)
 图 8 空间欧氏距离示意图

3 实验结果分析 3.1 实验条件

 图 9 CeleA部分人脸图像

3.2 算法测试 3.2.1 基于GrabCut的ROI提取实验

 图 10 图像预处理实验结果图

3.2.2 基于ResNet的人脸识别实验

 图 11 已知人脸数据库样本示意

 图 12 大样本人脸识别测试结果

 $P = \frac{{TP}}{{TP + FP}}$ (7)
 $R = \frac{{TP}}{{TP + FN}}$ (8)

3.3 系统测试

4 结论

 [1] He KM, Zhang XY, Ren SQ, et al. Deep residual learning for image recognition. Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, NV, USA. 2016. 770–778. [2] 张剑, 何骅, 詹小四, 等. 结合特征适配与拉普拉斯形变的3维人脸重建. 中国图象图形学报, 2014, 19(9): 1349-1359. DOI:10.11834/jig.20140912 [3] Neumann L, Matas J. Real-time scene text localization and recognition. Proceedings of 2012 IEEE Conference on Computer Vision and Pattern Recognition. Providence, RI, USA. 2012. 3538–3545. [4] Akbani O, Gokrani A, Quresh M, et al. Character recognition in natural scene images. Proceedings of 2015 International Conference on Information and Communication Techno-logies. Karachi, Pakistan. 2015. 1–6. [5] 慕春雷. 基于HOG特征的人脸识别系统研究[硕士学位论文]. 成都: 电子科技大学, 2013. [6] Lowe DG. Distinctive image features from scale-invariant key-points. International Journal of Computer Vision, 2004, 60(2): 91-110. DOI:10.1023/B:VISI.0000029664.99615.94 [7] Chen D, Cao XD, Wen F, et al. Blessing of dimensionality: High-dimensional feature and its efficient compression for face verification. Proceedings of 2013 IEEE Conference on Computer Vision and Pattern Recognition. Portland, OR, USA. 2013. 3025–3032. [8] Juefei-Xu F, Boddeti VN, Savvides M. Local binary convolutional neural networks. Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, HI, USA. 2017. 4284–4293. [9] Hasani B, Mahoor MH. Spatio-temporal facial expression recognition using convolutional neural networks and conditional random fields. Proceedings of the 12th IEEE International Conference on Automatic Face & Gesture Recognition. Washington, DC, USA. 2017. 790–795. [10] Krizhevsky A, Sutskever I, Hinton GE. ImageNet classification with deep convolutional neural networks. Proceedings of the 25th International Conference on Neural Information Processing Systems. Lake Tahoe, NV, USA. 2012. 1097–1105. [11] Szegedy C, Liu W, Jia YQ, et al. Going deeper with convolutions. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Boston, MA, USA. 2015. 1–9. [12] Simonyan K, Zisserman A. Very deep convolutional networks for large-scale image recognition. Proceedings of 2015 International Conference on Learning Representations. San Diego, CA, USA. 2015. 1482–1496. [13] Sun Y, Wang XG, Tang XO. Hybrid deep learning for face verification. Proceedings of 2013 IEEE International Conference on Computer Vision. Sydney, NSW, Australia. 2013. 1489–1496. [14] Taigman Y, Yang M, Ranzato M, et al. DeepFace: Closing the gap to human-level performance in face verification. Proceedings of 2014 IEEE International Conference on Computer Vision and Pattern Recognition. Columbus, OH, USA. 2014. 1701–1708. [15] 袁姮, 王志宏, 姜文涛. 刚性区域特征点的3维人脸识别. 中国图象图形学报, 2017, 22(1): 49-57. DOI:10.11834/jig.20170106 [16] 梁华刚, 易生, 茹锋. 结合像素模式和特征点模式的实时表情识别. 中国图象图形学报, 2017, 22(12): 1737-1749. DOI:10.11834/jig.170251 [17] Yu WS, Hou ZQ, Wang P, et al. Weakly supervised foreground segmentation based on superpixel grouping. IEEE Access, 2018, 6: 12269-12279. DOI:10.1109/ACCESS.2018.2810210 [18] Tang M, Gorelick L, Veksler O, et al. GrabCut in one cut. Proceedings of 2013 IEEE International Conference on Computer Vision. Sydney, NSW, Australia. 2013. 1769–1776. [19] 冯建文, 董剑. 改进的TCP应用层协议在远程实验系统中的应用. 计算机应用与软件, 2018, 35(12): 154-158.