Abstract: Aiming at the problems of low detection accuracy and high missed detection rate of small targets in the mask wearing detection algorithm in complex scenes, a mask wearing detection model based on improved YOLOV5s is proposed. In order to alleviate the problem of information loss caused by continuous downsampling, an inverted adaptive attention module（IAAM） is proposed. In order to improve the sensing ability of high-resolution detection layer to global information, a channel feature grouping module（CFGM） is designed, and the small object detection accuracy is greatly improved by using this module. Combined with the data features in the actual scene, EIoU Loss is introduced, and the color space transformation in data enhancement is cancelled. Experimental results show that the improved model has improved recall rate, detection accuracy, inference speed and detection ability of small targets, and can complete real-time mask wearing detection task in complex scenes.

Key words: mask wearing detection, YOLOV5s, small object detection, attention mechanism, feature grouping, EIoU loss

摘要： 针对复杂场景下口罩佩戴检测算法存在小目标检测精度低和漏检率高的问题，提出一种基于改进YOLOV5s的口罩佩戴检测模型。融合倒置自适应注意力模块IAAM（inverted adaptive attention module），缓解连续下采样导致的信息丢失问题，同时增强网络的特征融合能力；为了平衡多尺度检测层接收到的有效信息量，设计了通道特征分组模块CFGM（channel feature grouping module），提高了小目标检测精度；结合实际场景中的数据特征，使用EIoU Loss损失函数并取消数据增强中的色彩空间变换。实验结果表明，改进后的模型在检测精度、推理速度和小目标检测能力等方面均有提升，能够完成复杂场景下实时口罩佩戴检测任务。

关键词: 口罩佩戴检测, YOLOV5s, 小目标检测, 注意力机制, 特征分组, EIoU损失