摘要
由于信息技术的快速发展,红外检测技术和视频监控系统得到了广泛应用,图像型火灾探测器在火灾探测中的优势逐渐凸显。在飞机货舱火灾探测领域,尽管基于图像识别的火灾探测技术已展现出一定潜力,但在准确性与响应速度之间的平衡仍须进一步优化。为提升飞机货舱早期火灾的识别与判断能力,增强红外火焰图像目标检测的准确性,本文提出了一种结合改进损失函数的YOLO目标检测算法。首先对比了多种典型目标检测算法在红外火焰图像检测任务中的性能表现,进而选择了合适的算法框架进行损失函数的改进。通过在计算损失时综合考虑目标中心点的距离、重叠面积及长宽比等因素,设计了一种改进的损失函数,并成功将基于动态非单调聚焦机制的边界框损失(WIoU)函数引入YOLO目标检测网络中,以实现检测准确率的提升。训练红外火焰图像数据集的对比试验表明,改进后的YOLOv5算法在性能上并未取得显著提升,而YOLOv7算法在引入改进损失函数后,其检测精度相较于原算法提高了2.1%,平均精度值(mAP)提升了6.5%,同时每秒传输帧数(FPS)也增加了2.68帧。在关键的性能指标(如目标边框损失、置信度损失及总损失)上,采用WIoU损失函数的YOLOv7模型优于其他模型,达到了最低损失值。因此,本文提出的基于改进损失函数的YOLOv7算法在飞机模拟货舱红外火焰图像检测识别任务中展现出了更高的准确性和更快的处理速度,为飞机货舱火灾探测提供了一种有效的技术途径。
Owing to the swift evolution of information technology,infrared detection techniques and video surveillance systems have witnessed extensive utilization,wherein image-based fire detectors are increasingly proving their merits in fire detection.In the realm of aircraft cargo compartment fire detection,despite the demonstrated potential of imagebased fire detection technology,the equilibrium between precision and responsiveness necessitates further refinement.To bolster the capacity to identify and evaluate incipient fires within aircraft cargo compartments and augment the precision of infrared flame image target detection,this study introduces a refined YOLO(You Only Look Once)target detection algorithm integrated with an enhanced loss function.Initially,a comparison was made on the performance of multiple typical target detection algorithms in infrared flame image detection tasks,leading to the selection of a suitable algorithmic framework for the improvement of the loss function.Initially,a comparison was made on the performance of multiple typical target detection algorithms in infrared flame image detection tasks,leading to the selection of a suitable algorithmic framework for the improvement of the loss function.By meticulously accounting for variables like the distance between target centers,overlap area,and aspect ratio during loss computation,we crafted an enhanced loss function and effectively incorporated the weighted intersection over union(WIoU)loss function based on dynamic nonmonotonic focusing mechanism into the YOLO target detection network thereby bolstering detection accuracy.Experimental evaluations on infrared flame image datasets indicate that the enhanced YOLOv5 algorithm did not yield substantial gains in performance,whereas the YOLOv7 algorithm,after the introduction of the enhanced loss function,exhibited a 2.1%surge in detection accuracy,a 6.5%enhancement in mean average precision(mAP),and a 2.68-frame boost in frames per second(FPS).With regard to crucial performance metrics like box loss,objectness loss,and total loss,the YOLOv7 model utilizing the WIoU loss function excelled over other models,attaining the minimum loss value.In summary,the YOLOv7 algorithm with an enhanced loss function presented in this research demonstrates superior accuracy and responsiveness,offering a potent technical approach for aircraft cargo compartment fire detection.
作者
邓力
谢爽爽
刘全义
谭阳
Deng Li;Xie Shuangshuang;Liu Quanyi;Tan Yang(Civil Aircraft Fire Science and Safety Engineering Key Laboratory of Sichuan Province,Civil Aviation Flight University of China,Guanghan 618307,China;Sichuan Key Technology Engineering Research Center for All-electric Navigable Aircraft,Guanghan 618307,China)
出处
《航空科学技术》
2024年第11期112-118,共7页
Aeronautical Science & Technology
基金
国家自然科学基金(U2033206)
航空科学基金(20200046117001)
四川省重点实验室项目(MZ2022JB01)。
