摘要
在城市火灾发生时,社交媒体平台上的火灾现场照片对于火灾应急响应人员和决策者来说具有非常高的价值,可使决策者的决策制定更加高效,分配救援资源更加合理.然而,因社交媒体平台上图片数量巨大且伴随着大量与火灾现场不相关的火焰图片,所以需要采用一种方法对不相关图片进行筛选,且要保证速度与精确度。为此,文中提出了一种基于YOLOv5s网络构建对社交媒体平台图片实时筛选分类的方法,首先,通过对社交媒体平台上的火灾图像进行爬取,并对爬取到的3000张火灾图像数据集进行标注并构建数据集;然后,使用四种不同深度与宽度的YOLOv5s网络模型进行训练和测试;最后,对四种模型进行了比较和分析。实验结果表明,使用YOLOv5s模型训练效果整体优于其他三种模型,通过数据集测试网络模型能达到98.9%的精确率和92.1%的召回率以及96.9%的平均精度,检测每张图片的时间为0.009 s,很好地满足了实时筛选分类的要求。
When urban fires occur,fire scene photos on social media platforms are of great value to fire emergency responders and decision makers,helping them make more efficient decision making and allocate rescue resources more rationally.However,due to the large number of pictures on social media platforms,accompanied by a large number of flame pictures irrelevant to the fire scene,a method is needed to screen irrelevant pictures with high speed and accuracy.Therefore,a method of real-time filtering and classification of images on social media platforms is proposed based on Yolov5 s network construction.Firstly,the fire images on social media platforms are climbed,and 3000 fire image data sets are annotated and constructed.Then four YOLOv5 network models with different depth and width are used for training and testing.Finally,the four models are compared and analyzed.Experimental results show that the training effect of YOLOv5 s model is better than the other three models.Through the data set test network model can reach 98.9%accuracy,92.1%recall rate and 96.9%average accuracy,and the detection speed can reach 0.009 seconds,which well meets the requirements of real-time screening and classification.
作者
杨文阳
吴叶森
张峰
李湘眷
YANG Wen-yang;WU Ye-sen;ZHANG Feng;LI Xiang-juan(Xi'an Shiyou University,Xi'an 710065,China)
出处
《中国电子科学研究院学报》
北大核心
2022年第9期833-841,共9页
Journal of China Academy of Electronics and Information Technology
基金
国家自然科学基金(41301480)。
