摘要
2010—2022年,古建筑火灾事故频繁发生,对珍贵历史文物的保护提出了巨大挑战。为应对这一挑战,该研究提出一种基于图像超分辨率技术的消防危险品检测模型,用于识别易燃物品和可燃物品。在实际应用中,应准确归类危险品,并精确定位危险品的位置,由于普通摄像头图像的分辨率均较低,因而采用图像超分辨率生成对抗网络(super-resolution generative adversarial network,SRGAN)算法重建图像;此外,引入坐标注意力机制提高检测精度,并使用完整交并比(complete intersection over union,CIoU)损失函数加速模型训练。这一综合方法可有效检测古建筑区域内的消防危险品。使用测试数据集对该模型进行了一系列实验,结果表明:改进后的模型具有0.830的平均精度均值(mean average precision,mAP)和0.940的精确率;在服务器上,该模型的推理速度高达92 fps。该文提出模型有效且高效,为古建筑火灾防控工作提供了强大的技术支持。
[Objective] Ancient buildings are not just architectural structures but valuable cultural heritage sites that,when destroyed by fires,cause irreplaceable losses to history.In many instances,the root cause of these fires is not identified,although fire-hazardous items(such as cigarettes) often emerge as possible culprits.At present,fire risk assessment methods often fall short,offering subjective results and low detection rates that do not often meet the needs of practical application requirements.Therefore,we adopt a approach:real-time detection of fire-hazardous items such as cigarettes for effective monitoring and early warning.We aim to establish a model that provides accurate detection in real time using existing monitoring systems and video images.Considerably,we also plan a method for reconstructing blurred images.[Methods] To detect fire-hazardous items,five categories are selected:lighters,matches,candles,cigarettes,and incense.We independently create a dataset in VOC format,including 7 157 and 6 536 images of hazardous and non-hazardous items.Regarding the model,we choose the you only look once v5(YOLOv5) model as the starting point and made two key enhancements.First,we improve the generalized intersection over union(GIoU) loss function used in the YOLOv5 model by replacing it with complete intersection over union(CIoU),a new loss function.Unlike GIoU,CIoU considers the distance of the center point and aspect ratio,providing an accurate assessment of the quality of the predicted bounding box and leading to enhanced target box regression.Second,we tackle the issue of mismatched weight distribution between feature maps and channels by incorporating coordinate attention(CA) into the YOLOv5 backbone.CA reduces the number of channels in the feature map to increase the receptive field,learns the weight distribution of channels,and reallocates channel features based on these parameters.Ultimately,to reconstruct images,we utilize a super-resolution reconstruction algorithm based on the generative adversarial network,known as super-resolution generative adversarial network(SRGAN).SRGAN effectively removes blur from the original image,resulting in a more natural-looking reconstructed image and improving the overall quality of the dataset images.[Results] We implemented the ADAM optimizer to train the fire hazard detection model using a batch size of 16 for over 100 epochs.SRGAN-YOLOv5 performed exceptionally well in key metrics such as precision and recall,achieving impressive scores of 0.940 and 0.770,respectively.In terms of average precision for individual targets,we observed scores of 0.839,0.743,0.949,0.851,and 0.767 for lighters,cigarettes,candles,incense,and matches,respectively.This resulted in a mean average precision(mAP) of 0.830 across all categories.This model was tested against mainstream object detection networks for comparison,including faster regions-convolutional neural network(faster-RCNN),YOLOv3,YOLOv4,YOLOv5,and YOLOX.SRGAN-YOLOv5 ranked second in precision and mAP but boasted the highest frames per second(fps) rate.Considering the overall performance,SRGAN-YOLOv5 demonstrated significant advantages and extensive prospects in practical applications.In this paper,we also provided visualizations of the detection results achieved by the SRGAN-YOLOv5 model.[Conclusions] This work has led to the creation of a custom-built fire hazard dataset,can detect fire-hazardous items effectively.Utilizing SRGAN for image reconstruction,we successfully enhance the image resolution,which in turn improved the accuracy of the SRGAN-YOLOv5 model.As for the model itself,we refine the loss function and integrated the CA,leading to further enhancements in the precision of the SRGAN-YOLOv5 model.As a result,this model can achieve rapid and high-precision detection,making it a valuable tool for fire hazard detection.
作者
李悦明
黄国忠
王波
高学鸿
孙占辉
潘睿
LI Yueming;HUANG Guozhong;WANG Bo;GAO Xuehong;SUN Zhanhui;PAN Rui(Research Institute of Macro-Safety Science,University of Science and Technology Beijing,Beijing 100083,China;Institute of Public Safety Research,Department of Engineering Physics,Tsinghua University,Beijing 100084,China)
出处
《清华大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2024年第11期1870-1879,共10页
Journal of Tsinghua University(Science and Technology)
基金
国家重点研发计划项目(2021YFC1523502)。
作者简介
李悦明(1983-),男,博士研究生;通信作者:高学鸿,副教授,E-mail:gaoxh2020@ustb.edu.cn。
