基于光流和卷积视觉Transformer的轻量级微表情识别

徐恺蔚,KHIZER BIN TALIBHafiz,曹衍龙,许源平,许志杰,宋景春

Lightweight micro-expression recognition based on optical flow and convolutional vision Transformer

Kaiwei XU,Hafiz KHIZER BIN TALIB,Yanlong CAO,Yuanping XU,Zhijie XU,Jingchun SONG

表 6 模型中各模块的消融实验结果

Tab.6 Ablation experimental results of each module in model

模型序号 注意力映射 方式 分类器 输入 Full SMIC CASME Ⅱ SAMM UF1 UAR UF1 UAR UF1 UAR UF1 UAR 1 Linear class token 0.883 9 0.882 9 0.818 9 0.820 8 0.957 4 0.957 4 0.882 8 0.867 8 2 Conv_BN class token 0.904 9 0.905 3 0.842 0 0.848 2 0.978 5 0.975 4 0.903 0 0.874 8 3 InnerBN_DWConv class token 0.907 1 0.906 2 0.848 5 0.852 9 0.978 9 0.982 0 0.903 1 0.871 4 4 InnerBN_DWConv SeqSoftmax-g 0.909 8 0.909 8 0.846 6 0.851 7 0.978 0 0.968 8 0.921 2 0.915 8 5 InnerBN_DWConv SeqSoftmax-x 0.910 2 0.910 2 0.851 2 0.854 7 0.985 8 0.985 8 0.908 0 0.893 6