Fig. 1. Comparison of test protocol of face recognition. (a) Closed-set face recognition; (b) open-set face recognition
Fig. 2. Comparison of softmax loss function. (a) Traditional softmax loss function; (b) improved softmax loss function
Fig. 3. Schematic of the proposed angular distance loss function
Fig. 4. Structure of densely connected networks
Fig. 5. Comparison of activation functions. (a) ReLU; (b) PReLU
Fig. 6. Integral structure of network
Fig. 7. Face recognition accuracy versus hyperparameter ω
Fig. 8. Test accuracy of LFW dataset for network structures with different layer numbers and different loss functions
Fig. 9. Test accuracy of LFW dataset for network structures with different layer numbers and widths
Fig. 10. Proposed implementation process
Loss function | Decision boundary |
---|
Original softmax loss | (W1-W2)x+b1-b2=0 | Modified softmax loss | (cosθ1-cosθ2)=0 | Angular distance loss | {cosθ1-cos[(1-ω)θ2]}=0 for class 1 | {cos[(1-ω)θ1]-cosθ2}=0 for class 2 |
|
Table 1. Comparison of classification boundaries of loss functions
Layer | Output size | DenseFace-42 | DenseFace-54 | DenseFace-78 | DenseFace-122 |
---|
Dense block 1 | 56×56 | ×4 | ×6 | ×6 | ×6 | Dense block 2 | 28×28 | ×5 | ×6 | ×12 | ×12 | Dense block 3 | 14×14 | ×5 | ×6 | ×12 | ×24 | Dense block 4 | 7×7 | ×4 | ×6 | ×6 | ×16 |
|
Table 2. Specific configuration of the dense connection structure
Net structure | Input size /pixel | Depth /layer | Parameter /106 |
---|
LeNet | 32×32×1 | 5 | 0.062 | AlexNet | 227×227×3 | 8 | 62.4 | VGGNet | 224×224×3 | 16 | 138.4 | GoogleNet | 224×224×3 | 22 | 5.3 | ResNet | 224×224×3 | 152 | 61.3 | DenseFace (width: 32) | 112×112×3 | 42 | 6.7 | 54 | 7.3 | 78 | 8.9 | 122 | 12.8 | DenseFace (width: 16) | 112×112×3 | 42 | 5.78 | 54 | 5.9 | 78 | 6.37 | 122 | 7.4 |
|
Table 3. Comparison of parameter quantities of several convolutional neural network models
Method | Dataset | Data amount /106 | Accuracy /% |
---|
DeepFace | LFW | 4 | 97.33 | FaceNet | LFW | 200 | 99.67 | Deep FR | LFW | 2.6 | 98.85 | DeepID2+ | LFW | 0.3 | 98.74 | Center Face | LFW | 0.7 | 99.31 | Softmax loss | CAISA-WebFace | 0.49 | 97.78 | Triplet loss | CAISA-WebFace | 0.49 | 98.65 | Center loss | CAISA-WebFace | 0.49 | 99.02 | L-softmax loss | CAISA-WebFace | 0.49 | 99.15 | Angular distance loss | CAISA-WebFace | 0.49 | 99.45 |
|
Table 4. Test accuracy of different loss functions or face recognition algorithms
Method | Test protocol | Accuracy /% |
---|
Face identification | Face verification |
---|
FaceNet | large | 70.496 | 86.473 | Deepsense | large | 74.798 | 87.764 | Deepsense | small | 70.983 | 82.851 | Softmax loss | small | 54.628 | 65.732 | Triplet loss | small | 64.698 | 78.030 | Center loss | small | 65.334 | 80.106 | L-softmax loss | small | 67.035 | 80.185 | Angular softmax loss | small | 72.534 | 85.348 |
|
Table 5. Test accuracy of different loss functions or face recognition algorithms on the MegaFace dataset