摘要 在人群密集的场合下,佩戴口罩可以最大程度上防止病毒的传播。针对佩戴口罩检测任务常常会被复杂光照干扰这一问题,该文提出了基于YOLOv4改进的佩戴口罩检测算法。在YOLOv4模型的基础上,首先在主干特征提取网络中引入双注意力机制模型使模型对特征有更好的提取效果,之后在特征金字塔部分加入跨阶段局部网络,通过改进网络结构来减少参数量,在保证准确率的前提下,减小模型尺寸,提升推理速度。最后提出了一种图像筛选算法用来选取满足复杂光照这一条件的图像来制作数据集。实验结果表明,针对光照复杂多变的场合下对人脸佩戴口罩检测的平均精度均值可达到92.1%。与主流的目标检测算法相比,具有更好的检测效果。 In crowded situations,the virus can be prevented to the greatest extent possible by wearing a mask.Aiming at the problem that mask-wearing detection tasks are often interfered by complex illumination,an improved mask detection algorithm based on YOLOv4 is proposed.based on the YOLOv4 model,First,the dual attention mechanism model is introduced into the backbone feature extraction network to make the model have a better extraction effect on features,and then add a cross stage partial network to the feature pyramid.The network structure is improved to reduce the number of parameters,and under the premise of ensuring accuracy,the size of the model is reduced,and the inference speed is improved.Finally,an image filtering algorithm is proposed to select the images that meet the complex lighting conditions to make the data set.Experimental results show that the mean average precision of face mask detection can reach 92.1%under complex and changeable illumination.Compared with mainstream target detection algorithms,it has better detection results.
机构地区 青岛大学自动化学院
出处 《自动化与仪表》 2021年第4期67-72,78,共7页 Automation & Instrumentation
