基于属性驱动损失函数的人脸识别算法 下载: 757次
Face Recognition Algorithm Based on Attribute-Driven Loss Function
天津大学电气自动化与信息工程学院, 天津 300072
图 & 表
图 1. 使用A-Softmax损失函数训练的特征分布和使用属性驱动损失函数训练的特征分布比较
Fig. 1. Comparison of feature distributions trained by A-Softmax loss function and attribute-driven loss function
下载图片 查看原文
图 2. 超参数λ和η对验证准确率的影响。(a) η相同λ不同的验证准确率;(b) η不同λ相同的验证准确率
Fig. 2. Influences of super parameters λ and η on verification accuracy. (a) Verification accuracy with same η and different λ; (b) verification accuracy with different η and same λ
下载图片 查看原文
表 1二分类任务下不同损失函数的决策边界的比较
Table1. Comparison of decision boundaries of different loss functions in binary case
Loss function | Decision boundary |
---|
Softmax loss | (W1-W2)x+b1-b2=0 | Modified Softmax loss | ‖x‖cosθ1-‖x‖cosθ2=0 | A-Softmax | class 1: ‖x‖[cos(mθ1)-cosθ2]=0class 2: ‖x‖[cosθ1-cos(mθ2)]=0 |
|
查看原文
表 2不同网络结构的验证准确率比较
Table2. Comparison of verification accuracy of different network structures
Network structure | LFW /% | CFP-FP /% | AgeDB-30 /% |
---|
ResNet50 | 99.27 | 91.34 | 94.31 | ResNet101 | 99.35 | 92.15 | 95.82 | MobileNet | 99.13 | 90.10 | 93.88 | Inception-ResNet v2 | 99.67 | 93.00 | 97.42 | DenseNet | 99.54 | 92.39 | 96.40 | SE-ResNet101 | 99.48 | 92.78 | 96.67 |
|
查看原文
表 3不同损失函数的验证准确率
Table3. Verification accuracy of different loss functions
Loss function type | LFW /% | CFP-FP /% | AgeDB-30 /% |
---|
Softmax | 97.78 | 89.64 | 93.04 | Triplet loss | 98.65 | 90.22 | 95.88 | Center loss | 99.02 | 91.10 | 96.12 | L-Softmax loss | 99.15 | 91.90 | 96.20 | A-Softmax loss | 99.42 | 92.80 | 96.83 | Modified A-Softmax loss | 99.67 | 93.00 | 97.42 |
|
查看原文
表 4MegaFace数据集下不同损失函数的准确率
Table4. Accuracy of different loss functions in MegaFace dataset
Method | Protocol | Identificationaccuracy /% | Verificationaccuracy /% |
---|
Vocord-DeepVo1 | Large | 75.127 | 67.318 | Google-FaceNet V8 | Large | 70.496 | 86.493 | Softmax loss | Small | 54.628 | 65.732 | Triplet loss | Small | 64.698 | 78.030 | Center loss | Small | 65.334 | 80.106 | L-Softmax | Small | 67.035 | 80.185 | A-Softmax | Small | 72.729 | 85.561 | Modified A-Softmax loss | Small | 74.531 | 87.134 |
|
查看原文
李燊, 苏寒松, 刘高华, 吴慧华, 王萌. 基于属性驱动损失函数的人脸识别算法[J]. 激光与光电子学进展, 2019, 56(24): 241505. Shen Li, Hansong Su, Gaohua Liu, Huihua Wu, Meng Wang. Face Recognition Algorithm Based on Attribute-Driven Loss Function[J]. Laser & Optoelectronics Progress, 2019, 56(24): 241505.