基于改进SSD的口罩佩戴检测算法  

Mask Wearing Detection Algorithm Based on Improved SSD

作　　者：李雨阳 沈记全[1] 翟海霞[1] 冯伟华[2] LI Yuyang;SHEN Jiquan;ZHAI Haixia;FENG Weihua(College of Computer Science and Technology,Henan Polytechnic University,Jiaozuo,Henan 454000,China;Zhengzhou Tobacco Research Institute,China National Tobacco Corporation,Zhengzhou 450000,China)

机构地区：[1]河南理工大学计算机科学与技术学院,河南焦作454000 [2]中国烟草总公司郑州烟草研究院,郑州450000

出　　处：《计算机工程》2022年第8期173-179,186,共8页Computer Engineering

基　　金：国家自然科学基金(61972134);河南省科技攻关项目(182102310946)。

摘　　要：新冠疫情近年来在全球肆虐,新冠病毒具有极强的传染性,在公共场所佩戴口罩可以阻断病毒的传播途径,有效遏止疫情的蔓延。利用计算机视觉技术对公共场合佩戴口罩行为进行检测具有重要意义,在疫情防控常态化条件下,需要对人脸口罩佩戴进行正确识别,同时要识别口罩佩戴是否正确。在实际检测环境中,口罩佩戴检测任务中的场景复杂多样,佩戴口罩的人脸目标尺度不一,正确与错误佩戴口罩特征差异小,难以检测。提出一种改进SSD算法的口罩佩戴检测算法。以SSD检测算法为基础,引入特征融合网络及协调注意力机制,重构特征提取网络,增强对细节信息的学习和处理能力。同时,将算法的分类预测分数和IoU分数进行合并表示,使用Quality Focal Loss函数调节正负样本的权重。在自制口罩佩戴检测数据集上的实验结果表明,该算法的平均精度均值达到96.28%,与原始算法相比提高了5.62%,对口罩佩戴检测具有良好的准确性和实用性,可满足疫情防控下的实际需求。In recent years,the COVID-19,which involves a highly infectious virus,has spread worldwide.Wearing masks in public areas can reduce the transmission and hence the spread of the virus.Additionally,using computer vision technology to detect mask wearing behavior in public areas is crucial.To prevent and control epidemics,the correct form of wearing face masks must be identified.In an actual environment,the detection of mask wearing is complex and diverse.The scale of a face wearing a mask is different;furthermore,the difference between the correct and wrong forms of wearing a mask is subtle and hence difficult to detect.Therefore,a mask wearing detection algorithm based on an improved Single Shot Multibox Detector(SSD)algorithm is proposed herein.Based on the SSD network,the algorithm introduces a feature fusion network and an attention coordination mechanism,reconstructs the feature extraction network,and enhances the ability of learning and processing detailed information.In addition,the classification prediction score and IoU score of the algorithm are combined,whereas the Quality Focal Loss(QFL)function is used to adjust the weight of positive and negative samples.An experiment is performed on acustom-developed mask wearing test dataset.Experimental results show that the average accuracy of the algorithm is 96.28%,which is 5.62%higher than that of the original algorithm.The improved algorithm offers good accuracy and practicability for mask wearing detection,as well assatisfies the requirements for epidemic prevention and control.

关 键 词：口罩佩戴检测 SDD检测算法 特征融合 注意力机制 损失函数 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]