基于YOLOv4算法的口罩佩戴检测

MASK DETECTION BASED ON YOLOV4

  • 摘要: 针对公共场所口罩佩戴检测所面临的实时性、多目标多姿态以及面部遮挡等挑战,设计一种基于YOLOv4的口罩佩戴检测方法。该方法一方面通过引入空洞卷积和非对称卷积等思想,结合RFB设计特征加强模块RFB-s,并替换YOLOv4中的空间金字塔结构,扩大了模型感受野,同时降低了网络参数量。另一方面,增加了注意力模块,提升了模型信息处理能力。通过在自建的口罩佩戴检测数据集和开源数据集上的实验, 对比不同网络结构和不同算法情况下的mAP值和运行速度, 验证了该算法在口罩佩戴检测性能上的提升。

     

    Abstract: Mask detection in practical applications is facing the challenges of real-time requirement, multi-target, multi-pose, facial occlusion and complex scene. In this paper, an effective mask detection method based on YOLOv4 is proposed to deal with such challenging factors. In this method, a new neck block RFB-s was designed base on receptive field block (RFB) which replaced the spatial pyramid structure in YOLOv4 and introduced the dilated convolution and asymmetric convolution to expand the receptive field without parameter increase. In addition, an attention module was added to the prediction process to improve the information processing capability of the detection model. Experiments on mask detection dataset and open-source dataset verify the effectiveness of the proposed mask detection method by comparing mAP and running speeds under different network structures and algorithm.

     

/

返回文章
返回