多尺度特征融合的轻量化口罩佩戴检测算法

Lightweight Mask-Wearing Detection Algorithm with Multi-Scale Feature Fusion

科学规范地佩戴口罩是预防新冠、流感等呼吸道传染病的有效方法,在当前疫情形势下,正确佩戴口罩显得尤为重要。已有的口罩佩戴检测算法多数存在结构复杂、训练难度较高和特征提取不足等问题,为此,提出一种多尺度特征融合的轻量化口罩佩戴检测算法L-MFFN-YOLO。以YOLOv4-Tiny网络为基础,L-MFFN-YOLO改进原始残差结构,使用轻量化残差模块促进模型快速收敛,在有效降低模型计算量的同时保证检测精度。在原网络13×13、26×26这2个尺度的基础上增加52×52特征分支,以增强低特征层的信息表达能力并降低小目标的漏检率。通过多层级交叉融合结构最大程度地提取有用信息,从而提高特征利用率。除佩戴和未佩戴口罩2种情况外,在数据集中新增口罩佩戴不正确的类别并进行手工标注,实验结果表明,L-MFFN-YOLO算法的模型大小仅为5.8 MB,较原始网络YOLOv4-Tiny,其模型规模减小76%,mAP提高5.25个百分点,CPU下的处理时间快14 ms,能在资源受限的设备中满足口罩佩戴检测任务对准确率和实时性的要求。

Standardized usage of face masks is effective as a non-pharmaceutical intervention to prevent the spread of infectious respiratory diseases,such as COVID-19 and influenza. In the current epidemic situation,wearing face masks correctly is especially important. Most existing mask-wearing detection algorithms involve problems such as complex structures,high training difficulty,and insufficient feature extraction.Therefore,this study proposes a lightweight maskwearing detection algorithm based on multi-scale feature fusion and the YOLOv4-Tiny network,called L-MFFN-YOLO.L-MFFN-YOLO improves on the original residual structure and uses a lightweight residual module to promote rapid convergence. Moreover,it reduces the computational load while ensuring detection accuracy. Based on the original network’s 13×13 and 26×26 feature maps,52×52 feature branches are added to enhance the ability of the lower feature layer to express information and reduce the false negative rate for small targets.On this basis,a Multi-level Cross Fusion(MCF) structure is used to maximally extract useful information so as to improve feature utilization. In addition to detecting mask-wearing,a category of masks worn incorrectly is added to the dataset and manually labeled. The experimental results show that the size of the proposed L-MFFN-YOLO model is only 5.8 MB,which is 76% smaller than that of the original YOLOv4-Tiny.Moreover,the mean Average Precision(mAP)of the proposed approach is 5.25percentage points higher,and its processing time is 14 ms faster on an equivalent CPU.These results demonstrate that the proposed approach can meet the requirements of accuracy and real-time operation in resource-constrained devices to detect faces wearing masks.