基于改进AlexNet的人脸表情识别 
下载： 1512次
Facial Expression Recognition Based on Improved AlexNet
昆明理工大学信息工程与自动化学院, 云南 昆明 650500
图 & 表
图 1. 改进AlexNet网络结构
Fig. 1. Improved AlexNet network structure
下载图片 查看原文
图 2. 多尺度卷积网络结构
Fig. 2. Multi-scale convolutional network structure
下载图片 查看原文
图 3. JAFFE数据库的7种人脸表情
Fig. 3. Seven facial expressions in the JAFFE database
下载图片 查看原文
图 4. CK+数据库的7种人脸表情
Fig. 4. Seven face expressions in the CK+ database
下载图片 查看原文
图 5. 数据预处理流程
Fig. 5. Data preprocessing process
下载图片 查看原文
表 1改进AlexNet网络的参数
Table1. Parameters of improved AlexNet network
| Layer | Net | Input | Convolution size | Stride | Padding | Output |
|---|
| 1-1 | Convolution1-1 | 48×48×1 | 1×1×96 | 1 | Same | 48×48×96 | | 1-2 | Convolution1-2 | 48×48×1 | 3×3×96 | 1 | Same | 48×48×96 | | 1-3 | Convolution1-3 | 48×48×1 | 5×5×96 | 1 | Same | 48×48×96 | | 2 | Concatenate | - | - | - | - | 48×48×288 | | 3 | Maxpooling | 48×48×288 | 2×2 | 2 | Valid | 24×24×288 | | 4 | Convolution | 24×24×288 | 5×5×256 | 1 | Same | 24×24×256 | | 5 | Maxpooling | 24×24×256 | 2×2 | 2 | Valid | 12×12×256 | | 6 | Convolution | 12×12×256 | 3×3×384 | 1 | Same | 12×12×384 | | 7 | Convolution | 12×12×384 | 3×3×384 | 1 | Same | 12×12×384 | | 8 | Convolution | 12×12×384 | 3×3×256 | 1 | Same | 12×12×256 | | 9 | Maxpooling | 12×12×256 | 2×2 | 2 | Valid | 6×6×256 | | 10 | GAP (global average pooling) | 24×24×288 | - | - | - | 288 | | 11 | GAP | 12×12×256 | - | - | - | 256 | | 12 | Flatten | 6×6×256 | - | - | - | 9216 | | 13 | Concatenate | - | - | - | - | 9760 | | 14 | FC (fully connected layer) | 6×6×256 | - | - | - | 4096 | | 15 | FC | 4096 | - | - | - | 4096 | | 16 | Softmax | 4096 | - | - | - | 7 |
|
查看原文
表 2实验结果对比
Table2. Comparison of experimental results%
| Algorithm | AlexNet | Em-AlexNet |
|---|
| Accuracy of CK+ | 85.60 | 94.25 | | Accuracy of JAFFE | 78.57 | 93.02 |
|
查看原文
表 3不同方法在CK+数据集的对比
Table3. Comparison of different methods in CK+ dataset%
| Method | Accuracy |
|---|
| CNN[5] | 81.50 | | C-LeNet5[22] | 83.74 | | Local gabor+RFLD+KNN[23] | 91.51 | | Multi_resolution feature CNN[24] | 92.10 | | Em-AlexNet | 94.25 |
|
查看原文
表 4不同方法在JAFFE数据集的对比
Table4. Comparison of different methods in JAFFE dataset%
| Method | Accuracy |
|---|
| PHOG+LBP+SVM[25] | 87.43 | | Local gabor+RFLD+KNN[23] | 89.67 | | LDN+SVM[26] | 90.60 | | Multi_resolution feature CNN[24] | 91.70 | | Em-AlexNet | 93.02 |
|
查看原文
表 5AlexNet和Em-AlexNet算法对CK+数据集中不同表情的识别准确率
Table5. Recognition accuracy of AlexNet and Em-AlexNet algorithms for different expressions in the CK+ dataset%
| Method | Surprise | Angry | Contempt | Disgust | Fear | Happy | Sad | Total |
|---|
| AlexNet | 96.0 | 60.0 | 75.0 | 100.0 | 70.0 | 80.0 | 85.0 | 85.6 | | Em-AlexNet | 97.5 | 96.0 | 70.0 | 96.9 | 91.7 | 100.0 | 80.0 | 94.3 |
|
查看原文
表 6AlexNet和Em-AlexNet算法对JAFFE数据集中不同表情的识别准确率
Table6. Recognition accuracy of AlexNet and Em-AlexNet algorithms for different expressions in the JAFFE dataset%
| Method | Surprise | Angry | Disgust | Fear | Happy | Neutral | Sad | Total |
|---|
| AlexNet | 87.5 | 62.5 | 87.5 | 75.0 | 75.0 | 100.0 | 62.5 | 78.6 | | Em-AlexNet | 100.0 | 100.0 | 100.0 | 100.0 | 100.0 | 100.0 | 66.7 | 93.0 |
|
查看原文
杨旭, 尚振宏. 基于改进AlexNet的人脸表情识别[J]. 激光与光电子学进展, 2020, 57(14): 141026. Xu Yang, Zhenhong Shang. Facial Expression Recognition Based on Improved AlexNet[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141026.
PDF全文
