基于YOLOv4的轻量化口罩佩戴检测模型设计

Design of lightweight mask wearing detection model based on YOLOv4

受呼吸传染疾病的影响,公共场所佩戴口罩已成为常态,通过机器检测替代人工检查可以大大节省人力资源。常用的基于神经网络的目标检测模型往往参数量和计算量大,不利于终端部署和成本控制。面对上述问题,该文在YOLOv4通用化目标检测算法基础上,提出一种轻量化口罩检测模型G-YOLOv4。引入Ghost module模块搭建特征提取网络,在降低模型参数量和计算量的同时,提高了模型的检测精度和检测速度;在特征融合网络中引入深度可分离卷积,进一步降低模型计算量,减少过度拟合;使用Mish激活函数作为特征提取网络的激活函数,优化了模型的收敛效果。实验结果表明,G-YOLOv4算法的mAP为93.21%,相比于YOLOv4算法提高率为5.52%;模型参数大小为11 488 785,相比于YOLOv4算法减少率为82.05%;模型的计算量为7.05 GFLOPs,相比于YOLOv4算法减少率为72.90%;模型的检测效率(FPS)达到了38.23,相比于YOLOv4算法提升率为204.14%。

Affected by respiratory infectious diseases,wearing masks has become the norm in many public places.Using machine inspection instead of manual inspection can greatly save human resources.Commonly used general target detection models based on deep neural networks often have a huge number of parameters,which is not conducive to terminal deployment and cost control.In response to these problems,this paper proposes a lightweight mask detection model G-YOLOv4 based on the improved general target detection model YOLOv4.Using GhostNet to decrease the parameters and computation of the discovery model,the network also increase the accuracy of the model discovery.The use of depth separable convolution instead of the general convolution in the original YOLOv4 further reduces the computational complexity of the model.Use the Mish activation function as the activation function of the GhostNet shallow network to optimize the convergence effect of the model.The experimental results show that the mAP of the optimized G-YOLOv4 algorithm reaches 93.21%,which has an increase rate of 5.52%compared with the original algorithm.The model parameter size is 11488785,which has a reduction rate of 82.05%compared with the original algorithm.The calculation amount of the model is 7.05 GFLOPs,which has a reduction rate of 72.8%compared with the original algorithm.The detection efficiency(FPS)of the model reaches 38.23,which has an increase rate of 204.14%compared with the original algorithm.