| 摘 要: 准确检测人员是否佩戴口罩对于保证食品生产环境的卫生、预防疾病的传播等具有重要意义。文章改进了YOLOv5算法的网络结构，首先基于双向特征金字塔网络(BiFPN)改进了YOLOv5网络的颈部(Neck)结构，然后使用K-means聚类算法对数据集进行聚类分析，提出YOLOv5_BM口罩人脸检测算法。在自制人脸口罩数据集上的测试结果表明，YOLOv5_BM算法的平均精度高达95.3%，相比YOLOv5网络提升了3.8%。在公开数据集上与其他经典的目标检测算法相比，YOLOv5_BM算法在性能方面也取得了提升，相比SSD算法，YOLOv5_BM算法的平均精度提高了4.4%；相比YOLOv3算法，YOLOv5_BM算法的平均精度提高了2.9%。 |
| 关键词: 口罩人脸检测；YOLOv5；双向特征金字塔网络；K-means聚类算法 |
|
中图分类号: TP751.1
文献标识码: A
|
| 基金项目: 国家自然科学基金(61502340)；天津市研究生科研创新项目(2020YJSB077)；天津理工大学教学基金(YB20-05). |
|
| An Mask Face Detection Method of Improved Deep Learning based on BiFPN |
|
YU Xiao, ZHANG Maosong, ZHOU Zijie
|
(Tianjin University of Technology, Tianjin 300384, China )
yx_ustb@163.com; 1798588432@qq.com; Zhou_Zijie@outlook.com
|
| Abstract: Accurately detecting whether staff wear masks or not is of great significance for ensuring the sanitation of food production environment and preventing the spread of diseases. This paper proposes a YOLOv5_BM mask face detection algorithm by improving the network structure of YOLOv5 algorithm. Firstly, the neck part structure of YOLOv5 network is improved based on Bi-directional Feature Pyramid Network (BiFPN); then K-means clustering algorithm is used to cluster the data set. The test results on the self-made face mask dataset show that the average accuracy of YOLOv5_BM algorithm is 95.3%, which is 3.8% higher than YOLOv5 network. Compared with other classical target detection algorithms on the public dataset, YOLOv5_BM algorithm has also improved its performance. Compared with SSD (Single Shot MultiBox Detector) algorithm, the average accuracy of YOLOv5_BM algorithm has been improved by 4.4%; compared with YOLOv3 algorithm, the average accuracy of YOLOv5_BM algorithm is improved by 2.9%. |
| Keywords: mask face detection; YOLOv5; BiFPN; K-means clustering algorithm |
