基于YOLOv5s+CANet的人脸口罩检测改进算法  

Improved Face Mask Detection Algorithm Based on YOLOv5s+CANet

作　　者：吴慧婕 赵刚[1,2] 胡送惠 王正水 WU Hui-jie;ZHAO Gang;HU Song-hui;WANG Zheng-shui(School of Mathematics and Information Science,Nanchang Hangkong University,Nanchang 330063,China;Key Laboratory of Nondestructive Testing(Ministry of Education),Nanchang Hangkong University,Nanchang 330063,China)

机构地区：[1]南昌航空大学数学与信息科学学院,南昌330063 [2]无损检测技术教育部重点实验室(南昌航空大学),南昌330063

出　　处：《南昌航空大学学报（自然科学版）》2022年第4期108-115,共8页Journal of Nanchang Hangkong University(Natural Sciences)

基　　金：南昌航空大学博士启动基金(EA201907279);南昌航空大学重点科研基地开放基金(EW202107216)。

摘　　要：为解决实际应用场景中人脸检测可能会出现的光照模糊、遮挡物不明确等一系列问题,提高检测的精确度,本文基于YOLOv5s算法,针对人脸口罩目标检测进行网络模型改进:在YOLOv5s的主干网络部分和颈部的不同网络层级处加入注意力机制,针对边界框回归任务,替换YOLOv5s模型的损失函数,加速收敛提高回归精度。实验结果表明,当在YOLOv5s的backbone部分的P5层和Neck部分的P4、P5层之后加入Coordinate注意力机制,并用SIoU_Loss替换原本损失函数后,改进后的YOLOv5s算法与基准模型相比,精度值提升了1.6%。In order to address a series of problems such as blurred lighting and unclear occlusions that may occur in face detection in practical application scenarios, and promote the detection accuracy, the network model for face mask object detection was engineered based on the YOLOv5s algorithm, and through the following strategies: Adding an attention mechanism to the backbone network part of YOLOv5s and different network levels of the neck;Replacing the loss function of the YOLOv5s model for the bounding box regression task to accelerate convergence and improve regression accuracy. The experimental results show that when the Coordinate attention mechanism is added after the P5 layer of the backbone part in YOLOv5s and the P4 and P5 layers of the Neck part, and the original loss function is replaced with SIoU Loss, the accuracy rendered by the engineered YOLOv5s algorithm is increased by 1.5% compared with that of the benchmark model.

关 键 词：目标检测 卷积神经网络 YOLOv5s 注意力机制 CANet 

分 类 号：O29[理学—应用数学]