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 4 不同网络结构的对比实验结果

Tab.4 Comparative experimental results of different network structures

模型序号 模型类型 卷积嵌入层 注意力头数 Transformer层数 Full $K_i^{{\mathrm{e}}}$ $S_i^{{\mathrm{e}}}$ $P_i^{{\mathrm{e}}}$ ${C_i}$ UF1 UAR 1 基准三阶段模型MiER-CvT 32, 5, 3 16, 3, 1 8, 0, 0 64, 192, 384 1, 3, 6 1, 2, 3 0.907 1 0.906 2 2 不同层数的三阶段模型 32, 5, 3 16, 3, 1 8, 0, 0 64, 192, 384 1, 3, 6 1, 2, 4 0.905 1 0.906 6 3 32, 5, 3 16, 3, 1 8, 0, 0 64, 192, 384 1, 3, 6 1, 2, 5 0.906 9 0.905 8 4 32, 5, 3 16, 3, 1 8, 0, 0 64, 192, 384 1, 3, 6 1, 3, 3 0.902 6 0.907 0 5 32, 5, 3 16, 3, 1 8, 0, 0 64, 192, 384 1, 3, 6 1, 3, 4 0.906 0 0.901 8 6 32, 5, 3 16, 3, 1 8, 0, 0 64, 192, 384 1, 3, 6 2, 2, 3 0.899 0 0.900 5 7 不同嵌入维度的三阶段模型 32, 5, 3 16, 3, 1 8, 0, 0 64, 192, 768 1, 3, 12 1, 2, 3 0.903 3 0.904 9 8 32, 5, 3 16, 3, 1 8, 0, 0 64, 384, 768 1, 6, 12 1, 2, 3 0.907 4 0.909 3 9 常规卷积三阶段模型 7, 3, 3 4, 2, 2 2, 1, 1 64, 192, 384 1, 3, 6 1, 2, 3 0.844 3 0.847 0 10 双阶段模型 32, 5 16, 3 8, 0 64, 192 1, 3 1, 2 0.892 7 0.894 0 11 单阶段模型 32 16 8 64 1 1 0.768 2 0.752 6