非受控环境下基于混合注意力机制的面部表情识别

Study of Facial Expression Recognition based on Hybrid Attention Mechanism in Uncontrolled Environment

针对非受控环境下的面部表情识别过程中特征提取不充分的问题,构建了一个加强特征提取的网络结构——3_M_S_MobileNetV2模型,该结构以MobileNetV2为基础网络,结合自主设计的混合注意力模块3_M_CBAM,并且将ReLU6激活函数更换为SiLU激活函数设计而成。在两个非受控环境下所采集的数据集RAF_DB和Fer2013上对设计的模型进行实验,识别准确率分别达到81.88%和65.65%,并且与现有的基于神经网络的识别方法相比,在RAF_DB数据集和Fer2013数据集上的面部表情识别准确率分别提高了1%~9%、0.5%~4%,证明了文中网络结构的有效性。

To address the issue of insufficient feature extraction during the facial expression recognition process in uncontrolled environments,this paper presents a strengthened network structure—the 3_M_S_MobileNetV2—to optimize feature extraction.It is designed on the MobileNetV2 as the basic network and in combination with a self designed hybrid attention module—3_M_CBAM,with the ReLU6 activation function replaced by the SiLU activation function.The proposed model is tested on the two datasets,RAF_DB and Fer2013,acquired in two uncontrolled environments,which shows that the facial expression recognition accuracy reaches 81.88%and 65.65%respectively and is improved by 1%~9%and 0.5%~4%respectively compared with existing recognition methods which are based on neural networks.The tests have verified the feasibility of the network structure.