摘要
在林业管理中,及时发现火灾并识别其规模对于安全防护和治理火灾至关重要。针对现有火灾检测算法存在的精度低、漏检误检和实时性不足等问题,提出一种无人机航拍图像下火灾实时检测算法——MDSYOLOv8。以YOLOv8为基线算法,将骨干网络第7层卷积模块和颈部网络卷积模块替换成动态蛇形卷积(DSConv),提高算法的特征提取性能,并强化算法对微小特征的学习能力;然后在颈部与检测头之间添加多维协作注意力机制(MCA),加强颈部特征融合,增强算法对小目标的检测能力,并抑制无关背景信息;最后使用SIoU损失函数替换原YOLOv8中的CIoU损失函数,加快算法的收敛速度和回归精度。实验结果表明,MDSYOLOv8在公开数据集KMU上对烟雾目标的检测精度mAP达到95.89%,相较于基线YOLOv8提高了3.33个百分点,具有卓越的检测性能。此外,本研究采集互联网上的无人机航拍火灾图像制作UFF(UAV field fire)数据集,主要对象为火焰和烟雾,包含森林和城市等火灾隐患可能发生场景。在自制数据集UFF上进行深度实验分析,MDSYOLOv8的检测精度达到93.98%,检测速度为54帧/s,并且能同时识别烟雾和火焰两种火灾场景中的主要目标,与主流目标检测方法相比,在检测精度和效率方面均展现出明显优势,更加契合航拍场景下的火灾检测应用。
In forestry management,timely detection and identification of wildfires are crucial for ensuring safety protection and effective fire control.With the advancement of deep learning technologies,numerous object detection algorithms have been applied to fire detection tasks.However,these algorithms still have some shortcomings.To overcome challenges such as low accuracy,missed detections,false alarms,and inadequate real-time performance in current wildfire detection algorithms,this study proposed a real-time wildfire detection algorithm for unmanned aerial vehicle(UAV)aerial images:MDSYOLOv8.Building upon YOLOv8 as the baseline algorithm,the seventh layer convolution module in the backbone network and the convolution module in the neck network were replaced with dynamic snake convolution(DSConv)to enhance the algorithm s feature extraction performance and strengthen its learning capability for minor features.Subsequently,a multidimensional cooperative attention(MCA)mechanism was added between the neck and the detection head to enhance neck feature fusion,improve the algorithm s detection capability for small targets,and suppress irrelevant background information.Finally,the scale-invariant intersection over union(SIoU)loss function was used to replace the original complete intersection over union(CIoU)loss function in YOLOv8 to accelerate the algorithm s convergence speed and regression accuracy.Experimental results demonstrated that MDSYOLOv8 achieved a detection accuracy mean average precision(mAP)of 95.89%for smoke targets on the public dataset KMU,which was a 3.33 percentage points improvement over the baseline YOLOv8,showcasing outstanding detection performance.The KMU dataset was relatively limited,containing only smoke targets,making it unable to validate the algorithm s detection performance for flames.Moreover,the lack of scene diversity in the KMU dataset hindered the algorithm s generalization training.Therefore,this study collected drone-captured wildfire images from the internet to create the UFF(UAV field fire)dataset,focusing on flames and smoke and encompassing potential fire hazard scenarios in forests,cities,and other locations.In-depth experimental analysis conducted on the self-made UFF dataset showed that MDSYOLOv8 achieved a detection accuracy of 93.98%and a detection speed of 54 frames per second.It was capable of simultaneously identifying the main targets in both smoke and flame scenarios.Compared to mainstream object detection methods,MDSYOLOv8 demonstrated significant advantages in terms of detection accuracy and efficiency,making it more suitable for wildfire detection applications in drone-captured scenes.
作者
郭纪良
刘莉
何建
GUO Jiliang;LIU Li;HE Jian(Shandong Business Institute,Yantai 264000,China;School of Information and Electrical Engineering,Ludong University,Yantai 264000,China;College of Engineering,China Agricultural University,Beijing 100083,China)
出处
《林业工程学报》
北大核心
2025年第2期111-122,共12页
Journal of Forestry Engineering
基金
国家自然科学基金(61903172)
烟台市智慧城市创新实验室科研项目(SDGP370600000202302000504)。
