Fig. 1. Overall structure of proposed model
Fig. 2. Dual-channel generator network structure
Fig. 3. Discriminator network structure
Fig. 4. Generation effect of frontal expression images on KDEF dataset
Fig. 5. Generation effect of frontal expression images on Multi-PIE dataset
Fig. 6. Confusion matrix of expression recognition under different angles on KDEF dataset. (a) 0° rotation; (b) 45° rotation; (c) 90° rotation
Fig. 7. Partial error samples on KDEF dataset
Fig. 8. Confusion matrix of expression recognition under different angles on Multi-PIE dataset. (a) 0° rotation; (b) 30° rotation; (c) 60° rotation; (d) 90° rotation
Fig. 9. Partial error samples on Multi-PIE dataset
Fig. 10. Comparison of network loss function changes. (a) TP-GAN discriminant loss; (b) TP-GAN adversarial loss; (c) proposed model discriminant loss; (d) proposed model adversarial loss
Fig. 11. Comparison of facial expression images generation effect after model ablation. (a) Profile; (b) without local path; (c) without Wasserstein; (d) without GP; (e) proposed method; (f) real frontal
Fig. 12. Comparison of facial expression recognition rate under different ablation results
Fig. 13. Comparison of frontalization generation effect. (a) Real profile; (b) proposed method; (c) TP-GAN
[11]; (d) CAPG-GAN
[9]; (e) FI-GAN
[23]; (f) MTDNN
[24]; (g) 3DMM
[25]; (h) real frontal
Input | Operator | Channel | SE | Stride | Repeat |
---|
1282×3 | Conv 3×3 | 16 | | 2 | | 642×16 | Bottleneck 3×3 | 16 | √ | 2 | | 322×16 | Bottleneck 3×3 | 24 | | 2 | | 162×24 | Bottleneck 5×5 | 40 | √ | 2 | | 82×40 | Bottleneck 5×5 | 40 | √ | 1 | ×2 | 82×40 | Bottleneck 5×5 | 48 | √ | 1 | ×2 | 82×48 | Bottleneck 5×5 | 96 | √ | 2 | | 42×96 | Bottleneck 3×3 | 96 | √ | 1 | ×2 | 42×96 | Conv 1×1 | 256 | √ | 1 | | 42×256 | Pooling 4×4 | | | 1 | | 12×256 | Conv 1×1 | 512 | | 1 | | 12×512 | Conv 1×1 | 7 | | 1 | |
|
Table 1. Structure of improved MobileNetV3
Angle /° | -90 | -45 | 0 | 45 | 90 |
---|
Accuracy /% | 83.88 | 87.76 | 93.27 | 88.37 | 82.65 |
|
Table 2. Multi-angle facial expression recognition rate on KDEF dataset
Angle /° | -90 | -60 | -30 | 0 | 30 | 60 | 90 |
---|
Accuracy /% | 81.83 | 86.00 | 88.67 | 92.17 | 89.17 | 85.50 | 81.33 |
|
Table 3. Multi-angle facial expression recognition rate on Multi-PIE dataset
Method | 30° | 60° | 90° |
---|
Without Lsym (k3=0) | 87.00 | 81.83 | 76.17 | Without Lip (k4=0) | 82.67 | 73.33 | 62.50 | Proposed method | 89.17 | 85.50 | 81.33 |
|
Table 4. Expression recognition rate under different loss functions on Multi-PIE dataset
Method | Accuracy /% | Parameter computation /MB |
---|
VGG-19[20] | 82.15 | 70.45 | ResNet-50[21] | 84.94 | 23.60 | Xception[22] | 90.55 | 20.88 | MobileNetV1[14] | 90.24 | 3.24 | Proposed method | 92.17 | 0.95 |
|
Table 5. Recognition results of frontal expressions under different models
Method | -90˚ | -45˚ | 0˚ | 45˚ | 90˚ |
---|
Reference[5] | 74.50 | 72.00 | 70.50 | 77.40 | 79.50 | Reference[6] | | 80.39 | 84.07 | 77.03 | | Reference[7] | | | 86.67 | 83.81 | 76.67 | Proposed method | 83.88 | 87.76 | 93.27 | 88.37 | 82.65 |
|
Table 6. Comparison of multi-angle facial expression recognition rate with existing methods on KDEF dataset
Method | -90˚ | -60˚ | -30˚ | 0˚ | 30˚ | 60˚ | 90˚ |
---|
Reference[26] | | | 93.10 | 92.80 | 92.20 | | | Reference[27] | | | 92.58 | 89.65 | 94.51 | | | Reference[28] | 79.00 | 83.00 | 80.00 | 86.00 | 83.50 | 82.00 | 76.00 | Reference[29] | | | | 88.20 | 87.30 | 83.80 | 78.80 | Reference[30] | | | | 83.30 | 85.80 | 81.60 | 92.50 | Proposed method | 81.83 | 86.00 | 88.67 | 92.17 | 89.17 | 85.50 | 81.33 |
|
Table 7. Comparison of multi-angle facial expression recognition rate with existing methods on Multi-PIE dataset