基于多运动特征融合的微表情识别算法 
下载： 1221次
Micro-Expression Recognition Algorithm Based on Multiple Motive Feature Fusion
天津大学电气自动化与信息工程学院, 天津 300072
图 & 表
图 1. 整体网络框架
Fig. 1. Overall network framework
下载图片 查看原文
图 2. LMF检测过程示意。(a)人脸中68个关键点的示例;(b) LMF可视化结果
Fig. 2. Diagram of LMF detection process. (a) Example of 68 monitoring points in face; (b) LMF visualization
下载图片 查看原文
图 3. 光流和光学应变示意图。(a)微表情开始帧;(b)微表情峰值帧;(c)光流可视化图;(d)光学应变可视化图
Fig. 3. Diagram of optical flow and optical strain. (a) Start frame of micro-expression; (b) peak frame of micro-expression;(c) optical flow visualization image; (d) optical strain visualization image
下载图片 查看原文
表 1在CNN模块上不同数据的性能对比
Table1. Performance comparison of different data in CNN module
| Input data | Accuracy /% |
|---|
| Optical strain | 47.98 | | LMF | 45.70 | | Optical flow | 41.50 | | Original data | 36.65 |
|
查看原文
表 2单一特征图和多特征融合图结果的对比
Table2. Result comparison of single feature map and multiple feature fusion map
| ExperimentNo. | Opticalstrain | LMF | Opticalflow | Accuracy /% |
|---|
| 1 | √ | | | 52.45 | | 2 | | √ | | 51.38 | | 3 | | | √ | 50.54 | | 4 | √ | √ | √ | 58.53 |
|
查看原文
表 3不同算法的各种指标对比
Table3. Comparison of indicators of different algorithms
| Algorithm | Accuracy /% | UAR /% | F1-score /% |
|---|
| LBP-TOP* | 45.95 | 30.94 | 29.41 | | LBP-SIPr | 46.56 | | 44.80 | | FDMo | 45.93 | | 40.53 | | ELRCN-TE* | 50.00 | 39.37 | 43.4 | | MRWo | 46.15 | | 43.07 | | STCLQPo | 58.39 | | 58.36 | | MDMOr | 44.25 | | 44.16 | | Bi-WOOFo | 57.89 | | 61.25 | | Proposed algorithm | 58.53 | 48.07 | 52.82 |
|
查看原文
苏育挺, 王蒙蒙, 刘婧, 费云鹏, 何旭. 基于多运动特征融合的微表情识别算法[J]. 激光与光电子学进展, 2020, 57(14): 141504. Yuting Su, Mengmeng Wang, Jing Liu, Yunpeng Fei, Xu He. Micro-Expression Recognition Algorithm Based on Multiple Motive Feature Fusion[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141504.
PDF全文
