基于图卷积神经网络的面部动作单元检测

Facial Action Unit Detection Based on Graph Convolutional Neural Network

情绪识别已广泛应用于教学效果评估和心理疾病检测等场景,面部动作单元检测是情绪识别的关键步骤。在图卷积神经网络基础上,融合残差网络(Residual Network,ResNet)、压缩激励网络(Squeeze and Excitation Networks,SENet)、全卷积神经网络(Fully Convolutional Networks,FCN)4种网络结构,建立带有注意力机制的面部动作单元检测模型,并在丹佛大学自发面部运动单元数据库(Denver Intensity of Spontaneous Facial Action,DISFA)和CK+两个公共数据集上进行了验证实验。实验结果表明,该模型的性能优于传统面部动作单元检测模型。

Emotion recognition has been widely applied in teaching effectiveness evaluation and psychological disease detection scenarios,and facial action unit detection is a key step in emotion recognition.On the basis of graph convolutional neural networks,a facial action unit detection model with attention mechanism is established by integrating four network structures:Residual Network(ResNet),Squeeze and Excitation Networks(SENet),and Fully Convolutional Networks(FCN),And validation experiments were conducted on the Denver Intensity of Spontaneous Facial Action(DISFA)and CK+public datasets at the University of Denver.The experimental results show that the performance of this model is superior to traditional facial action unit detection models.