餐饮从业人员应佩戴口罩是基本要求，但佩戴是否规范仍未引起充分关注。为此，针对油烟、水汽、火光等复杂后厨环境提出一种口罩佩戴规范性检测方法。首先采集并创建了具有针对性的数据集CKEMFD-12k；其次构建了一个多任务卷积神经网络MMWN来提取判断口罩佩戴情况所需的关键要素信息，通过自主设计的多尺度混联空间金字塔池化模块MHSPP和跨任务交互管式注意力机制TETAM，相比现有算法取得了94.68%的最高目标检测准确率、4.62%的最小口鼻关键点平均定位误差和94.32%的最高口罩区域分割准确率；最后计算口鼻关键三角区并解析其与口罩区域的覆盖关系，给出规范佩戴、不当佩戴和没有佩戴口罩的判定原理。实验表明，综合判断准确率为93.57%，超越现有主流算法。

The basic requirement for catering staff to wear masks is widely acknowledged, but it still has not paid sufficient attention to mask wearing specification. To deal with this, a detection method for mask wearing specification was proposed for complex working environments of kitchen such as oil fumes, water vapor, and flames. Firstly, a targeted dataset named CKEMFD-12k was collected and constructed. Secondly, a multi-task convolutional neural network(MMWN) was constructed to extract key element information for assessing mask wearing status. By utilizing the self-designed multi-scale hybrid spatial pyramid pool module(MHSPP) and tube-embedded transformer attention mechanism(TETAM), the method achieved the highest average target detection accuracy of 94.68%, the minimum mouth-nose key points mean error of 4.62%, and the optimal mask region segmentation pixel accuracy of 94.32% compared to the existing network. At last, a algorithm was designed to calculate mouth-nose key triangle area and analyze their coverage relationship with the mask area, which provides a judgment method or three mask wearing status: normative wearing, improper wearing, and no wearing. Experiment shows that the comprehensive judgment accuracy is 93.57%, surpassing existing mainstream algorithms.

TP391.4???????

福建省自然科学基金资助项目（2023J011439，2019J01859）