深度回归网络下的人脸对齐方法

Face alignment method in deep regression networks

为解决在人脸识别、表情识别等过程中,大规模姿态变化和遮挡等复杂环境下无法实现精确面部对齐的问题,提出一种深度回归网络(deep regression networks)下的人脸对齐方法。使用深度卷积神经网络对测试样本进行头部姿态估计并初始化人脸形状,降低头部姿态对人脸对齐造成的影响,在深度回归网络下进行局部回归微调,实现抗遮挡人脸精确对齐,使用图基的双权重函数作为训练深度回归网络的损失函数,减小训练过程中异常值的影响。选用300W和COFW人脸数据集进行测试,实验结果表明,该方法相比其它方法具有较强的鲁棒性,可实现大规模姿态变化和遮挡下人脸的精确对齐。

To solve face alignment in complex environment such as large-scale pose change and occlusion,in the process of face recognition and facial expression recognition,a face alignment method based on deep regression networks was put forward.The head pose of the test sample was estimated using deep convolutional neural networks and the facial shape was initialized,by which the effect of head pose on face alignment was reduced.Through the deep regression networks for local regression fine-tuning,anti-occlusion face accurate alignment was implemented.The Tukey＇s biweight function was used as the loss function of training deep regression networks,by which the influence of outliers in the training process was minimized.300 W and COFW face data sets were selected for testing.Experimental results show that the proposed method is more robust compared to other methods,and precise alignment of the face under large-scale pose change and occlusion was realized.