Face multi-attribute detection algorithm based on RetinaFace
-
摘要: 为提高铁路刷脸检票业务中人脸检测的平均精度,通过研究分析人脸检测算法RetinaFace,针对闸机应用场景制定损失函数,提出了一种基于RetinaFace的人脸多属性检测算法,实现了人脸框位置、人脸是否佩戴墨镜以及人脸遮挡程度等信息的准确输出。算法使用轻量化骨干网络MobileNet-0.25网络结构,移除非必要的分支,减少计算开销,在铁路标准人脸遮挡数据集上检出率达到95.4%,不同遮挡程度的识别准确率达到了99.2%。Abstract: In order to improve the average accuracy of face detection in railway face brushing and ticket checking, this paper proposed a face multi-attribute detection algorithm based on RetinaFace by studying and analyzing the face detection algorithm retinaface and formulating the loss function according to the application scene of the gate. It was implemented the accurate output of the information such as the face frame position, whether the face was wearing sunglasses, and the degree of occlusion. The algorithm used the lightweight backbone network MobileNet-0.25 network structure, and removed unnecessary branches to reduce the computational cost. The detection rate of the algorithm on the railway standard face occlusion data set reached 95.4%, and the recognition accuracy of different occlusion degrees reached 99.2%.
-
Key words:
- occlusion recognition /
- loss function /
- feature pyramid /
- convolution neural network /
- object detection
-
表 1 各个层级目标框的数目
特征金字塔 步长 锚点框 P2(160$ \times $160$ \times $256) 4 16,20.16,25.40 P3(80$ \times $80$ \times $256) 8 32,40.32,50.80 P4(40$ \times $40$ \times $256) 16 64,80.63,101.59 P5(20$ \times $20$ \times $256) 32 128,161.26,203.19 P6(10$ \times $10$ \times $256) 64 256,322.54,406.37 表 2 实验结果
模型 CPU时间 简单数据集
检出率困难数据集
检出率墨镜检出率 遮挡程度
准确率本文改进
模型35 ms 95.4% 89.7% 99.4% 99.2% RetinaFace 32 ms 95.3% 85.4% - - -
[1] 贾成强,戴琳琳,徐海涛,等. 基于人脸识别技术的铁路实名制进站核验系统研究及设计 [J]. 铁路计算机应用,2018,27(7):49-53, 63. doi: 10.3969/j.issn.1005-8451.2018.07.012 [2] Kaiming H, Georgia G, Piotr D, et al. Mask R-CNN [J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2020, 42(2): 386-397. [3] Redmon J, Divvala S, Girshick R, et al. You only look once: unified, real-time object detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Los Alamitos: IEEE Computer Society Press, 2016: 779-788. [4] W. Liu, D. Anguelov, D. Erhan, et al. SSD: single shot multibox detector[C]//Proceedings of the 14th European Conference onComputer Vision. Amsterdam, the Neterlands, 2016: 21-37. [5] Deng J, Guo J, Ververas E, et al. RetinaFace: Single-Shot Multi-Level Face Localisation in the Wild[C]//2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2020. [6] Lin T Y, Dollar P, Girshick R, et al. Feature Pyramid Networks for Object Detection[C]//IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, 2017. [7] K. He, X. Zhang, S. Ren, et al. Deep residual learning forimage recognition[C]//IEEE Conference on Computer Visionand Pattern Recognition (CVPR). Las Vegas, 2016: 770-778. [8] GlorotX, Bengio Y. Understanding the difficulty of training deep feedforward neural networks [J]. Journal of Machine Learning Research, 2010(9): 249-256. [9] Howard A G, Zhu M, Chen B, et al. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications [J]. IEEE Access, 2017(6): 1-14. [10] Yang S, Luo P, Loy C C, et al. WIDER FACE: A Face Detection Benchmark[C]//IEEE Conference on Computer Vision & Pattern Recognition. IEEE, 2016.