基于光流和集成时-空-通道注意力的ResNet-10的微表情识别模型

Micro-expression recognition model based on optical flow and integrated spatio-temporal-channel attention of ResNet-10.

针对一般模型很难捕捉微表情不同尺度上的特征,提出一种基于LiteFlowNet和改进的ResNet-10的微表情识别网络以充分提取微表情不同维度信息。先通过欧拉视频放大技术(EVM)突出面部微小动作,再将处理后的数据通过轻量级光流估计网络LiteFlowNet提取视频帧中的运动信息。在用于特征提取的ResNet-10上引入三维注意力机制(3D-Attention),以适应性地聚焦于微表情视频中最具辨别力的通道、空间和时间特征。实验结果验证了该网络有效提升了微表情识别性能。

In response to the difficulty of general models to capture the features of micro-expressions at different scales,a micro-expression recognition network based on LiteFlowNet and the improved ResNet-10 is proposed to fully extract the information of different dimensions of micro-expression.The facial micro-movements are first highlighted by EVM,and then the processed data are passed through a lightweight optical flow estimation network,LiteFlowNet,to extract the motion information in the video frames.3D-Attention mechanism is introduced on ResNet-10 for feature extraction to adaptively focus on the most discriminative channel,spatial and temporal features in the micro-expression video.The experimental results verify that the network effectively improves the micro-expression recognition performance.