一种基于注意力模型的面部表情识别算法 
下载： 1340次
An Attention Model-Based Facial Expression Recognition Algorithm
天津大学电气自动化与信息工程学院, 天津 300072
图 & 表
图 1. CSACNN模型结构
Fig. 1. CSACNN model structure
下载图片 查看原文
图 2. 通道注意力分支
Fig. 2. Channel attention branch
下载图片 查看原文
图 3. 空间注意力分支
Fig. 3. Spatial attention branch
下载图片 查看原文
图 4. 注意力模型集成与残差学习单元
Fig. 4. Attention model integration and residual learning unit
下载图片 查看原文
图 5. 面部68个关键点
Fig. 5. 68 face landmarks
下载图片 查看原文
图 6. 面部表情关键区域的截取。(a)原图;(b)面部遮罩;(c)截取图像
Fig. 6. Cropping of key areas of facial expression. (a) Original image; (b) facial mask; (c) cropped image
下载图片 查看原文
表 1超参数对网络性能的影响
Table1. Effect of hyper-parameters on network performance
| Variable | Value | Accuracy /% |
|---|
| 1 | 97.35 | | d | 4 | 97.45 | | 8 | 97.25 | | 8 | 95.72 | | r | 16 | 97.45 | | 32 | 95.41 |
|
查看原文
表 2注意力模型的位置对网络性能的影响
Table2. Effect of attention model location on network performance
| Dataset | Location | Accuracy /% |
|---|
| After conv | 96.64 | | CK+ | Before pooling | 97.45 | | After pooling | 95.72 | | After conv | 72.69 | | MMI | Before pooling | 74.73 | | After pooling | 72.59 |
|
查看原文
表 3不同表情识别方法性能对比
Table3. Performance comparison of different expression recognition methods
| Method | Experimentalsetting | Accuracy /% |
|---|
| CK+ | MMI |
|---|
| 3DCNN[32] | Sequence-based | 85.90 | 53.20 | | LBP-TOP[33] | Sequence-based | 88.99 | 59.51 | | HOG 3D[34] | Sequence-based | 91.44 | 60.89 | | STM-ExpLet[35] | Sequence-based | 94.19 | 75.12 | | DTGAN[18] | Sequence-based | 97.25 | - | | Island Loss[25] | Image-based | 94.39 | 74.68 | | IACNN[22] | Image-based | 95.37 | 71.55 | | DLPCNN[24] | Image-based | 95.78 | - | | DeRL[36] | Image-based | 97.30 | 73.23 | | Ref.[19] | Image-based | 97.37 | - | | PPDN[37] | Image-based | 97.30 | - | | VGG16(ours) | Image-based | 91.72 | 64.13 | | ResNet5(ours) | Image-based | 86.87 | 57.09 | | CSACNN(ours) | Image-based | 97.45 | 74.73 |
|
查看原文
表 4不同模块的性能对比
Table4. Performance comparison of different modules
| Model | Accuracy /% |
|---|
| CK+ | MMI |
|---|
| Base | 94.73 | 67.61 | | Base+CA | 95.21 | 71.47 | | Base+SA | 95.62 | 70.17 | | Base+Crop | 95.82 | 71.41 | | Base+CA+SA | 96.43 | 72.98 | | Base+CA+SA+Crop | 97.45 | 74.73 |
|
查看原文
褚晶辉, 汤文豪, 张姗, 吕卫. 一种基于注意力模型的面部表情识别算法[J]. 激光与光电子学进展, 2020, 57(12): 121015. Jinghui Chu, Wenhao Tang, Shan Zhang, Wei Lü. An Attention Model-Based Facial Expression Recognition Algorithm[J]. Laser & Optoelectronics Progress, 2020, 57(12): 121015.
PDF全文
