基于深度学习的人脸多属性识别及其并行算法研究与探索

Face attribute recognition is widely used in video surveillance, mobile payment and other fields. It is increasingly important to study the accurate real-time recognition of face attributes. With the gradual maturity of deep learning technology, convolution neural network and other algorithms have outperformed classical learning algorithms in face attribute recognition. However, most of the current algorithms are only applied to single attribute prediction, and can not simultaneously predict multiple attributes that are related to each other. They can not give full play to the advantages of the correlation between attributes, which can enhance the original feature information of each attribute and make it easier to be recognized. In view of this, this project first uses the deep neural network to mine the relationship among different attributes from the bottom, middle and high levels in a progressive way. On this basis, a three-dimensional data model of high-level features is constructed, and the tensor correlation analysis algorithm is used to enhance the original single attribute to realize the prediction of multiple face attributes. At the same time, aiming at the insufficiency of the model training and prediction process, a parallel computing model combining data parallelism and model parallelism is proposed. Finally, based on the multi-GPU cluster platform, fast recognition of multiple attributes is realized in the open datasets of face images. This project is of great significance to improve the accuracy and speed of face attribute recognition.

人脸属性识别在视频监控、移动支付等领域具有广泛应用，对人脸属性进行精确实时识别的研究日显重要。随着深度学习技术的逐渐成熟，卷积神经网络等算法在人脸属性识别方面的效果已优于经典学习算法。但是，目前多数算法仅限于应用于单个属性预测，无法同时对多个互相关联的多个属性进行预测，不能发挥属性间关联关系可增强每个属性原有特征信息从而更易识别的优点。鉴此，本项目首先拟采用深度神经网络从底层、中层以及高层以递进方式挖掘不同属性之间的关联关系，在此基础上，构建高层特征三维数据模型，利用张量关联分析算法对原有单个属性进行特征增强，以实现对多个人脸属性进行预测。同时，针对模型训练和预测过程耗时过长的不足，拟使用数据并行和模型并行相结合的并行计算模型，最后，基于多GPU集群平台，在人脸图片公开数据集上实现多属性的快速识别。本项目对提升人脸属性识别准确率和识别速度具有重要意义。

人脸识别广泛运用在我们日常生活中，例如监控系统、安防系统、银行系统等，为我们生活带来巨大方便。但人脸识别算法依旧受到光照、噪音、数据集大小、实时性需求等各方面严峻的挑战，如何提升人脸识别系统识别准确率，增加系统鲁棒性，并让系统具有良好实时性识别效果。本项目主要从以下几个方面展开研究：1）人脸多属性关联增强方法研究，对人脸性别、种族、年龄和笑容等多个属性之间的关联关系进行探索研究，设计一种人脸多属性关联增强网络用于人脸多属性的预测；2）人脸老化算法研究，主要预测未来当前人脸图像可能的老化图像；3）人脸识别系统鲁棒性测试和提升研究，设计相应的对抗攻击算法来测试人脸识别系统防御性，并将该方法用于提升人脸识别系统防御性。基于这些研究，目前成果发表在ACM TOMM，ACM TIST，IEEE TCSVT（2篇），IEEE IOT，以及ACM MM CCF A类会议上。首先将原有的人脸识别准确率从80%提升到95%左右，系统鲁棒性大大提升，实时性从原来的1ms降低为0.1ms。这些成果对实时人脸识别算法研究及提升人脸识别算法鲁棒性具有重要意义。