一种多任务面部特征点与头部姿态检测方法

A Multi-Task Facial Landmarks and Head Pose Detection Method

目前国内在面部特征点检测和头部姿态估计研究领域多采用单任务方式,即分别对两项任务建立独立模型进行检测处理,忽视了二者之间可共享的隐层特征。面部特征点是表达头部姿态的重要信息,二者之间有复杂的非线性映射关系,基于其间可共享的隐层特征使用多任务学习可使两任务相互促进,优化检测速度与准确率。本文证明了头部姿态与面部特征点同时变化时二者具有的强相关性,并据此设计了一种多任务卷积神经网络以将面部特征点检测和头部姿态估计两项任务关联于一个神经网络模型中,共享核心卷积神经网络提取到的特征,后采用独立的分类器进行检测,最终以5个面部特征点和3个头部姿态参数为目标输出。实验表明,相比于传统单任务独立检测方法,采用多任务卷积神经网络可以同时完成面部特征点检测与头部姿态估计两项任务,并且在检测速度、精度上有较大的提升。

At present,in the research fields of facial landmarks detection and head pose estimation,single-task methods are mostly adopted in China,and independent models are established for the two tasks to carry out corresponding detection and recognition process separately.A multi-task convolutional neural network(MTL-CNN)is designed,which associates the two tasks in one neural network model sharing the features extracted by the core convolutional neural network to prove that the two tasks of facial landmarks detection and head pose estimation have a strong correlation.Independent classifiers are used for two tasks,and finally 5 facial landmarks and 3 head pose parameters are the target output.Experiments show that compared to single-task detection methods,the use of multi-task convolutional neural networks can simultaneously complete the two tasks of facial landmarks detection and head pose estimation,and the detection speed and accuracy are greatly improved.