基于盒注意力机制和Transformer的人脸微表情识别方法

Facial Micro-expression Recognition Method Based on Box-Attention Mechanism and Transformer

微表情是一种细微的能够体现人真实心理活动的面部运动,通常与真实情感直接相关,应用前景广阔。但由于微表情持续时间短暂、表情幅度低和特征难以提取等特点,因此其识别准确率较低。针对该问题,提出了基于盒注意力机制和Transformer的人脸微表情识别模型(visiontransformerbasedonbox-attention,ViT-Box)。该模型首先对人脸面部进行特征提取,利用盒注意力机制获得自适应的面部微表情关键区域:左眉眼、右眉眼和嘴巴;然后对非关键区域进行掩码遮盖,避免微表情无关信息干扰;最后基于VisionTransformer网络实现人脸微表情识别。ViT-Box模型在微宏表情仓库(MMEW)数据集上取得了98.68%的平均准确率,实验结果表明该模型在微表情识别上能够获得优秀的识别效果。同时通过消融实验验证了ViT-Box模型的有效性。

Micro-expression is a kind of subtle facial movement that can reflect people s real psychological activities,which is usually directly related to real emotions and has wide application prospects.However,due to the short duration of micro-expression,low range of expression and difficult to extract features,the recognition accuracy is low.To solve this problem,a new face micro-expression recognition model,vision Transformer based on Box-attention(ViT-Box),is proposed in this paper.Firstly,the model extracts facial features,and uses the Box-attention mechanism to obtain the key areas of self-adaptive facial micro-expressions,such as left eyebrow,right eyebrow and mouth.Then mask the non-critical areas to avoid the interference of irrelevant micro-expression information.Finally,facial micro-expression recognition is realized based on vision Transformer network.The experimental results show that the average accuracy of ViT-Box model on Micro-macro expression warehouse(MMEW)dataset is 98.68%,which indicates that the model can obtain excellent recognition effect in micro-expression recognition.The effectiveness of the ViT-Box model was verified by ablation experiments.