Author Affiliations
1School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China2School of Marine Science and Technology, Tianjin University, Tianjin 300072, Chinashow less
Fig. 1. Log-Gabor magnitude features of local facial expression image
Fig. 2. Second-order HOG features of local facial expression image
Fig. 3. Structure of DBN
Fig. 4. Flowchart of facial expression recognition based on fusion of local features and DBN
Fig. 5. Sample images. (a) JAFFE database; (b) CK database; (c) CK+ database
Fig. 6. Examples of facial expression database image preprocessing. (a) JAFFE database; (b) CK database; (c) CK+ database
Fig. 7. Expression recognition rate of DBN with different RBM layers
Database | 1 RBMlayer | 2 RBMlayers | 3 RBMlayers | 4 RBMlayers |
---|
JAFFE | 344.86 | 228.90 | 265.38 | 711.62 | CK | 337.76 | 402.75 | 537.88 | 669.64 | CK+ | 369.32 | 461.98 | 542.23 | 743.40 |
|
Table 1. Training and recognition time of DBN with different RBM layers
Feature | JAFFEdatabase | CKdatabase | CK+database |
---|
Gabor | 87.96 | 90.20 | 88.42 | Log-Gabor | 93.52 | 94.77 | 93.29 | HOG | 85.19 | 88.24 | 86.83 | Secondorder HOG | 92.59 | 94.12 | 92.68 | Log-Gabor+Second order HOG | 96.30 | 97.39 | 95.73 |
|
Table 2. Recognition rate based on different features
Algorithm | JAFFEdatabase | CKdatabase | CK+database |
---|
KNN | 75.00 | 78.43 | 77.44 | SVM | 82.41 | 83.01 | 81.10 | DBN | 96.30 | 97.39 | 95.73 |
|
Table 3. Recognition rate of different algorithms%
Method | Recognition rate /% |
---|
PHOG+LBP+SVM[24] | 87.43 | Local Gabor+RFLD+KNN[25] | 89.67 | LDN+SVM[26] | 90.60 | HOG+bagging ELM[27] | 94.37 | Proposed method | 96.30 |
|
Table 4. Comparison of recognition rate of different methods on JAFFE database
Method | Recognition rate /% |
---|
Local Gabor+RFLD+KNN[25] | 91.51 | LBP+MTSL[28] | 91.53 | CLBP+SVM[29] | 94.20 | GLDPE[30] | 97.08 | Proposed method | 97.39 |
|
Table 5. Comparison of recognition rate of different methods on CK database
Method | Recognitionrate /% |
---|
Geometric features+LBP+SVM[31] | 90.08 | HOG+DBN+Gabor+SAE[19] | 91.11 | PHOG+LBP+SVM[24] | 94.63 | Boosted DBN[20] | 96.70 | Proposed method | 95.73 |
|
Table 6. Comparison of recognition rate of different methods on CK+ database