Author Affiliations
1 School of Communication & Information Engineering, Shanghai University, Shanghai 200444, China;2 Key Laboratory of Advanced Display and System Applications, Ministry of Education, Shanghai University, Shanghai 200072, Chinashow less
Fig. 1. Framework of face landmark location based on LCCDN model
Fig. 2. Structural diagram of LSTM global network
Fig. 3. Descent curves of CNN loss functions in different face regions
Fig. 4. Classification performance of multi-pose face recognition under different Smax. (a) TPR; (b) FPR
Fig. 5. Examples of some clips in surveillance video dataset
Fig. 6. Qualitative results of facial landmark location based on LCCDN with various poses
Fig. 7. Accuracies of different face recognition methods with various poses
Network layer | Type | Filter | Output sizes | Others |
---|
Input | Input | - | 40×40 | - | Convolution 1 | Convolution | 5×5×20 | 36×36×20 | - | Maxpooling 1 | Max-pooling | 2×2 | 18×18×20 | - | Convolution 2 | Convolution | 3×3×40 | 16×16×40 | - | Maxpooling 2 | Max-pooling | 2×2 | 8×8×40 | - | Convolution 3 | Convolution | 3×3×60 | 6×6×60 | - | Maxpooling 3 | Max-pooling | 2×2 | 3×3×60 | - | Unshared Conv | Convolution | 2×2×80 | 2×2×80 | - | Fully-connectedfc1 | Fully-connected | - | 120 | - | Dropout1 | Dropout | - | 120 | Keep_ratiois 0.6 | Predicttion | Fully-connected | - | - | - |
|
Table 1. Structure of shallow CNN
Method | Mean error /10-2 | Failure rate /% |
---|
ERT[3] | 7.96 | 13.06 | AAM[4] | 7.58 | 12.56 | CFCNN[9] | 6.31 | 10.20 | TCDCN[15] | 4.60 | 6.59 | LCCDN | 4.06 | 5.26 |
|
Table 2. Experimental comparison of different facial landmark location methods
Facial orientationdescriptor | Accuracy /% |
---|
CASPEAL-R1[12] | CFP[13] | Multi-PIE[14] |
---|
D-LGBPH [16] | 91.75 | 91.25 | 92.63 | ASIFT[17] | 91.53 | 90.77 | 91.03 | CFCNN[9] | 93.54 | 92.28 | 94.16 | TCDCN[15] | 93.89 | 94.24 | 94.82 | LCCDN | 96.75 | 96.50 | 97.82 |
|
Table 3. Experimental comparison of multi-pose face recognition based on different face orientation descriptors
Clusteringmethod | Accuracy /% |
---|
CASPEAL-R1[12] | CFP[13] | Multi-PIE[14] |
---|
BART[11] | 93.85 | 94.13 | 95.28 | FCM[18] | 90.47 | 91.28. | 90.86 | K-means[19] | 91.39 | 92.05 | 93.90 | CBART | 96.75 | 96.50 | 97.82 |
|
Table 4. Experimental comparison of multi-pose face recognition based on different clustering methods
Method | Accuracy /% |
---|
CASPEAL-R1[12] | CFP[13] | Multi-PIE[14] |
---|
HPN[2] | 90.15 | 89.17 | 89.39 | VGGFace[7] | 93.20 | 92.89 | 92.78 | TPCNN [20] | 90.89 | 90.53 | 91.39 | DFLP[21] | 92.56 | 91.25 | 92.16 | Proposed | 96.75 | 96.50 | 97.82 |
|
Table 5. Experimental comparison of different face recognition methods