人脸面部微表情是一种快速而微妙的变化动作,为了更好地提取到微表情的运动特征,提出一种基于差分图像序列的微表情特征提取算法,可以有效地提取到微表情帧序列与起始帧之间的运动特征.为了提高微表情识别的准确率,提出了一种融合通道注意力和空间注意力的混合注意力模型,该模型可以很好地提取到微表情的关键特征,并且减少无用特征对微表情情绪识别结果的影响.将微表情的光流特征和差分图像序列特征分为两路送入混合注意力机制模型中,对模型进行训练并得到最终识别结果.在CASMEⅡ、SMIC(HS)、SAMM和MEGC2019数据集上的实验结果表明,使用双路特征比单一特征未加权平均召回率(Unweighted Average Recall,UAR)和未加权F4得分(Unweighted F1-score,UF1)均提升了 2％以上,使用混合注意力机制比不使用注意力机制的UAR和UF1均提升了 10％以上.
Micro-expression Recognition Based on Dual Path Features and Mixed Attention
Facial micro-expressions are fast and subtle changing movements.In order to better extract the motion features of micro-expressions,a micro-expression feature extraction algorithm based on differential image sequences is proposed,which can effectively extract the motion features between the micro-expression frame sequence and the starting frame.In order to improve the accuracy of micro-expression recognition,a hybrid attention model integrating channel attention and spatial attention is proposed,which can well extract the key features of micro-expressions and reduce the influence of useless features on the micro-expression emotion recognition results.The optical flow features of micro-expressions and differential image sequence features are fed into the hybrid attention mechanism model in two channels,and the model is trained and the final recognition results are obtained.The experimental results on CASME Ⅱ,SMIC(HS),SAMM and MEGC2019 datasets show that Unweighted Average Recall(UAR)and Unweighted F1-score(UF1)is improved more than 2％by using two-channel features than single features,and UAR and UF1 is improved more than 10％by using hybrid attention mechanism than without attention mechanism.
万方数据
维普
