Abstract: Helmet wearing recognition is a target detection task with less classification. The existing large-scale deep learning network model with high accuracy is used to identify helmet wearing, which has problems of parameter redundancy and large calculation, which is not suitable for deployment in embedded devices with limited computation to adapt to the actual site environment. To solve these problems, a lightweight network model YOLO-Ghost-BiFPNs3 suitable for embedded devices is proposed.?On the basis of YOLOv4, a new network structure is reconstructed based on Ghost module, and the depth and width of the network are trimmed. BiFPNs3, a lightweight module based on weighted channel addition, is designed to replace the FPN+PAN structure which has a large amount of calculation. A more quantifiable H-Swish activation function is used. Experiments are carried out on the Safety-Helmet-Wearing-Dataset dataset. In the test set, mAP@0.5 is 91.1%, which only loses 1 percentage point of accuracy compared with YOLOv4, and is 26 percentage points higher than the lightweight network model YOLOv4-Tiny. The number of parameters is 3% of that of YOLOv4, and the computational amount is only 5.8% of that of YOLOv4.

Key words: target detection, lightweight network model, helmet wearing identification, Ghost module

摘要： 安全帽佩戴识别是一种分类少的目标检测任务，使用现有精度较高的大型深度学习网络模型来识别安全帽佩戴，存在参数冗余问题且计算较大，不利于部署在计算量有限的嵌入式设备中以适应实际的工地环境。针对以上问题，提出了一种适合部署在嵌入式设备中的轻量化网络模型YOLO-Ghost-BiFPNs3。在YOLOv4的基础上，基于Ghost模块重构新的网络结构并对网络的深度和宽度进行裁剪；设计一种基于通道加权相加的轻量化模块BiFPNs3来替换原来计算量较大的FPN+PAN的结构；采用更容易量化的H-Swish激活函数；在Safety-Helmet-Wearing-Dataset数据集上进行实验，在测试集上，mAP@0.5为91.1%，相较于YOLOv4精度仅损失1个百分点，比轻量化网络模型YOLOv4-Tiny精度高26个百分点。参数量为原来YOLOv4的3%，计算量仅为原来YOLOv4的5.8%。

关键词: 目标检测, 轻量化网络模型, 安全帽佩戴识别, Ghost模块