深圳大学学报理工版

身份特征与表情特征是人脸图像分析中的两组重要特征,传统的有监督正交人脸特征学习(supervised orthogonal facial feature learning, SOFFL)算法虽然能够在给定表情和身份标签时学习这一对特征,但因数据要求较高令其应用受限.提出一种低数据要求的无监督正交人脸特征学习(unsupervised orthogonal facial feature learning, UOFFL)算法,通过提取正交人脸特征的统一框架,假设人脸图像空间中仅有身份和表情变化,使用重构损失、分类损失和相关性最小化损失的组合,采用深度卷积-反卷积神经网络,从已对齐的人脸图像中联合学习,提取身份和表情特征.其中,分类损失用于学习表情特征; 相关性最小化损失用于提高身份特征和表情特征之间的独立性; 重构损失用于确保两组特征组合的信息完整性.在大规模合成人脸表情数据集(large-scale synthesized facial expression dataset, LSFED)和受限的Radboud人脸数据集(Radboud faces dataset, RaFD)上进行验证,将所学身份特征空间中的欧氏距离用于人脸验证任务,结果表明,算法性能接近联合贝叶斯等有监督人脸识别方法.UOFFL算法可在身份标签缺失的条件下,仅使用表情特征学得身份特征.相比改进前的SOFFL算法, 该方法缓解了对身份标签的依赖, 适用场合更广.

In facial image analysis, the identity and expression are two important features which are related to face recognition and facial expression recognition tasks. The traditional supervised orthogonal facial feature learning method can learn both features jointly once the emotion and identity labels are given. Its application scope is limited by the high data requirement. In this paper, we propose a new method which is a united framework of joint facial feature learning with lower data requirement. It is based on the assumption that there are only two variations in the face space, i.e., the identity and the emotion. By combining reconstruction loss, classification loss, and correlation minimization loss, we use the deep convolutional-deconvolutional neural network to learn and extract identity and expression features from the aligned facial image. The classification loss learns the expression feature. The correlation minimization loss keeps the two features independent of each other. The reconstruction loss confirms the information completeness of the face. When the identity label is missing, the proposed method can learn the identity feature based on the expression labels. By this means, the new method is a kind of unsupervised facial feature learning. It relieves the limitation of the existing method, extends the application scope. The proposed method is evaluated on the large-scale synthesized facial expression dataset and the constrained Radbound face dataset(RaFD)dataset. The Euclid distance in the identity feature space is used for facial recognition task in which the performance of the proposed unsupervised facial feature learning is close to some supervision ones including the joint Bayesian face method. Our method relieves the requirement of identity label in the previous supervised version to get a wider application scope.

引言
1 基于深度特征分解的人脸特征学习
2 无监督正交人脸特征学习
3 实验与结果分析
4 结语

图1 SOFFL的网络结构[1]<br/>Fig.1 The network architectures of SOFFL[1]

图1 SOFFL的网络结构[1]
Fig.1 The network architectures of SOFFL[1]

图2 UOFFL的网络结构<br/>Fig.2 The network architectures of UOFFL

图2 UOFFL的网络结构
Fig.2 The network architectures of UOFFL

表1 UOFFL的网络细节<br/>Table 1 Details of the network for UOFFL

表1 UOFFL的网络细节
Table 1 Details of the network for UOFFL

图3 部分预处理后的LSFED与RaFD数据集人脸<br/>Fig.3 Some preprocessed faces in LSFED dataset and RaFD dataset

图3 部分预处理后的LSFED与RaFD数据集人脸
Fig.3 Some preprocessed faces in LSFED dataset and RaFD dataset

表2 基于不同空间欧氏距离的无监督人脸识别性能结果1)<br/>Table 2 Unsupervised face recognition results based on Euclidean distance in different spaces

表2 基于不同空间欧氏距离的无监督人脸识别性能结果1)
Table 2 Unsupervised face recognition results based on Euclidean distance in different spaces

表3 有监督和无监督人脸验证方法的性能对比1)<br/>Table 3 Comparisons among supervised and unsupervised facial verification methods

表3 有监督和无监督人脸验证方法的性能对比1)
Table 3 Comparisons among supervised and unsupervised facial verification methods

[1] SUN Wenyun, ZHAO Haitao, JIN Zhong. A complementary facial representation extracting method based on deep learning[J]. Neurocomputing, 2018, 306(6): 246-259.
[2] LANGNER O, DOTSCH R, BIJLSTRA G, et al. Presentation and validation of the Radboud faces database[J]. Cognition and Emotion, 2010, 24(8): 1377-1388.
[3] TRAN L, YIN Xi, LIU Xiaoming. Disentangled representation learning gan for pose-invariant face recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, USA: IEEE, 2017: 1415-1424.
[4] ZHANG Zhifei, SONG Yang, QI Hairong. Age progression/regression by conditional adversarial autoencoder[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Honolulu, USA: IEEE, 2017: 5810-5818.
[5] MA Liqian, SUN Qianru, GEORGOULIS S, et al. Disentangled person image generation[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018: 99-108.
[6] BERTHELOT D, RAFFEL C, ROY A, et al. Understanding and improving interpolation in autoencoders via an adversarial regularizer[EB/OL].(2018- 07-18)[2018- 07-23]. https://arxiv.org/abs/1807.07543, 2018.
[7] ZHU Zhenyao, LUO Ping, WANG Xiaogang, et al. Multi-view perceptron: a deep model for learning face identity and view representations[C]// Proceedings of the 27th Conference on Neural Information Processing Systems. Montreal, Canada: MIT Press, 2014, 1: 217-225.
[8] DOSOVITSKIY A, SPRINGENBERG J T, BROX T. Learning to generate chairs with convolutional neural networks[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Boston, USA: IEEE, 2015: 1538-1546.
[9] SUN Baochen, FENG Jiashi, SAENKO K. Return of frustratingly easy domain adaptation[C]// Proceedings of the 30th AAAI Conference on Artificial Intelligence. Phoenix, USA: AIII Press, 2016: 2058-2065.
[10] SUN Baochen, SAENKO K. Deep coral: correlation alignment for deep domain adaptation[C]// Proceedings of the European Conference on Computer Vision.[S. l.]: Spring, 2016: 443- 450.
[11] BOUSMALIS K, TRIGEORGIS G, SILBERMAN N, et al. Domain separation networks[C]// Advances in Neural Information Processing Systems. New York, USA: Curran Associates Inc., 2016: 343-351.
[12] GATYS L A, ECKER A S, BETHGE M. Image style transfer using convolutional neural networks[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas, USA: IEEE, 2016: 2414-2423.
[13] ROZANTSEV A, SALZMANN M, FUA P. Beyond sharing weights for deep domain adaptation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 41(4): 801-814.
[14] LI Yanghao, WANG Naiyan, LIU Jiaying, et al. Demystifying neural style transfer[C]// Proceedings of the International Joint Conference on Artificial Intelligence. Palo Alto, USA: AAAI Press, 2017: 2230-2236.
[15] LIAO Qianli, LEIBO Joel Z, POGGIO Tomaso. Learning invariant representations and applications to face verification[C]// Advances in Neural Information Processing Systems. Lake Tahoe, USA: MIT Press, 2013: 3057-3065.
[16] IOFFE S, SZEGEDY C. Batch normalization: accelerating deep network training by reducing internal covariate shift[EB/OL].(2015- 02-11)[2015- 03- 02]. https://arxiv.org/abs/1502.03167, 2015.
[17] KAZEMI V, SULLIVAN J. One millisecond face alignment with an ensemble of regression trees[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Columbus, USA: IEEE, 2014: 1867-1874.
[18] HASSNER T, HAREL S, PAZ E, et al. Effective face frontalization in unconstrained images[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Boston, USA: IEEE, 2015: 4295- 4304.
[19] HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Delving deep into rectifiers: surpassing human-level performance on imageNet classification[C]// Proceedings of the IEEE International Conference on Computer Vision. Santiago, Chile: IEEE, 2015: 1026-1034.
[20] CHEN Dong, CAO Xudong, WEN Fang, et al. Blessing of dimensionality: high-dimensional feature and its efficient compression for face verification[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Portland, USA: IEEE, 2013: 3025-3032.
[21] DONAHUE J, JIA Yangqing, VINYALS O, et al. DeCAF: a deep convolutional activation feature for generic visual recognition[C]// Proceedings of the International Conference on Machine Learning. Beijing: JMLR.org, 2014, 32: 647- 655.

备注

引言

1 基于深度特征分解的人脸特征学习

2 无监督正交人脸特征学习

3 实验与结果分析

4 结语

期刊信息

备注

引言

1 基于深度特征分解的人脸特征学习

2 无监督正交人脸特征学习

3 实验与结果分析

4 结 语

期刊信息

4 结语