融合梯度特征的轻量级神经网络的人脸识别 
下载： 773次
Face Recognition Based on Lightweight Neural Network Combining Gradient Features
1 东北石油大学电子科学学院, 黑龙江 大庆 163318
2 黑龙江省高校校企共建测试计量技术及仪器仪表工程研发中心, 黑龙江 大庆163318
图 & 表
图 1. Squeeze层与Expand层的降维流程图
Fig. 1. Flowchart of dimension reduction of Squeeze layer and Expand layer
下载图片 查看原文
图 2. 原始SqueezeNet模型流程图与改进后的流程图
Fig. 2. Flowchart of traditional SqueezeNet model and improved model
下载图片 查看原文
图 3. 图像分块为8×8时融合一阶梯度特征在LFW数据集上的训练模型收敛图
Fig. 3. Convergence diagram of training model of the proposed algorithm on LFW dataset for face images with 8×8 blocks
下载图片 查看原文
图 4. ROC曲线图
Fig. 4. Diagram of ROC
下载图片 查看原文
表 1算法在人脸图像不同分块大小的人脸识别率
Table1. Face recognition rate of the different algorithms for face images with different block sizes%
| Algorithm | No block | Block 2×2 | Block 4×4 | Block 8×8 | Block 16×16 |
|---|
| SqueezeNet | 91.63 | 92.10 | 92.17 | 92.92 | 91.19 | | SqueezeNet+first-step gradient feature | 91.99 | 92.65 | 93.19 | 97.28 | 96.97 | | SqueezeNet+second-step gradient feature | 92.62 | 93.47 | 95.70 | 98.33 | 97.39 |
|
查看原文
表 2算法在人脸图像不同分块大小的匹配时间
Table2. Matching time of the different algorithms for face images with different block sizess
| Algorithm | No block | Block 2×2 | Block 4×4 | Block 8×8 | Block 16×16 |
|---|
| SqueezeNet | 20.98 | 22.15 | 30.22 | 39.24 | 50.86 | | SqueezeNet+first-step gradient feature | 22.16 | 25.49 | 39.72 | 50.38 | 64.73 | | SqueezeNet+second-step gradient feature | 97.02 | 173.15 | 308.83 | 480.97 | 3047.31 |
|
查看原文
表 3人脸图像分8×8块时各个数据集下各种算法的识别率
Table3. Recognition rate of the different algorithms based on each data set with face images of 8×8 blocks%
| Algorithm | LFW | CASIA-WebFace | ORL |
|---|
| SqueezeNet | 92. 92 | 91.92 | 95.49 | | SqueezeNet+first-step gradient feature | 97.28 | 96.58 | 97.27 | | SqueezeNet+second-step gradient feature | 98.33 | 97.72 | 98.96 |
|
查看原文
表 4人脸图像分8×8块时各个数据集下各种算法的匹配时间
Table4. Matching time of the different algorithms based on each data set with face images of 8×8 blockss
| Algorithm | LFW | CASIA-WebFace | ORL |
|---|
| SqueezeNet | 39.24 | 23.17 | 9.02 | | SqueezeNet+first-step gradient feature | 50.38 | 37.65 | 27.96 | | SqueezeNet+second-step gradient feature | 480.97 | 279.69 | 169.46 |
|
查看原文
表 5算法在不同分块大小下交叉验证的识别率
Table5. Face recognition rate of cross validation algorithm for face images with different block sizes%
| Algorithm | No block | Block 2×2 | Block 4×4 | Block 8×8 | Block 16×16 |
|---|
| SqueezeNet | 76.33 | 76.75 | 77.87 | 80.04 | 79.15 | | SqueezeNet+first-step gradient feature | 76.98 | 77.23 | 78.69 | 80.95 | 79.63 |
|
查看原文
刘祥楼, 李天昊, 张明. 融合梯度特征的轻量级神经网络的人脸识别[J]. 激光与光电子学进展, 2020, 57(16): 161005. Xianglou Liu, Tianhao Li, Ming Zhang. Face Recognition Based on Lightweight Neural Network Combining Gradient Features[J]. Laser & Optoelectronics Progress, 2020, 57(16): 161005.
PDF全文
