基于多尺度优化感知网络的口罩检测方法

A mask detection method based on multi-scale optimized awareness network

佩戴口罩是全球医学专家公认最有效的预防新冠肺炎感染的方法之一。基于视觉的智能口罩检测技术对于督促人们在公共场合佩戴口罩具有重要的作用。然而,目前专用口罩检测算法较为缺乏,通用目标检测算法对于多尺度、多角度和外观多样的戴口罩人脸目标识别仍然无法满足检测精度的要求。针对该问题,提出了一种基于多尺度优化感知网络的口罩检测方法——PyramidMask。首先,Pyramid-Mask从骨干网络的不同尺度获取图像的多层特征;然后设计尺度感知分支进行不同层的高密度先验框独立预测,以端到端的方式对图像中多尺度的人脸进行精准定位和佩戴口罩检测。此外,为了提高模型对复杂环境的鲁棒性,在训练阶段以图像拼接的方式对训练样本进行数据增强。实验结果表明,在公开的口罩检测数据集上,PyramidMask优于当前主流方法。相较于基准方法,PyramidMask在检测戴与未戴口罩的召回率上分别有5.4%和12.5%的提升,精确率上分别有6.0%和4.1%的提升。

Wearing a mask is recognized by global medical experts as one of the most effective ways to prevent COVID-19 infection.The vision-based intelligent mask detection plays an important role in urging people to wear masks in public.However,compared with general object detection,there are currently few studies focusing on mask detection.To solve the problem,an optimized multi-scale awareness network,called PyramidMask,is proposed for mask detection.Firstly,PyramidMask obtains the multi-layer features of the image from different scales of the backbone.Secondly,the scale-awareness branches are designed to perform independent predictions of different layers of high-density candidate boxes.Finally,the multi-scale faces with masks in an image is accurately detected in an end-to-end manner.In addition,in order to improve the robustness of PyramidMask under complex scenes,the training samples are augmented by image stitching in the training stage.The experimental results show that PyramidMask outperforms the state-of-the-art methods on the public mask detection dataset.Compared with the benchmark,PyramidMask improves 5.4%and 12.5%in the recall of detection with and without masks,and 6.0%and 4.1%in the precision of detection with and without masks.