改进YOLOv8的航拍图像森林火灾检测算法  

Improved YOLOV8 Forest Fire Detection Algorithmin Aerial Images

作　　者：陈林豪 李广明 申京傲 邹永钶 欧阳裕荣 CHEN Linhao;LI Guangming;SHEN Jing’ao;ZOU Yongke;OUYANG Yurong(School of Computer Science and Technology,Dongguan University of Technology,Dongguan 523808,China)

机构地区：[1]东莞理工学院计算机科学与技术学院,广东东莞523808

出　　处：《东莞理工学院学报》2025年第1期48-56,共9页Journal of Dongguan University of Technology

基　　金：广东省自然科学基金(2022A1515140119,2023A1515011307)。

摘　　要：在全球气候变化背景下,森林火灾对环境和公共安全的威胁日益增加。针对无人机火灾预警技术在长距离航拍视角下存在的多尺度及多形变目标的误检和漏检问题,提出一种改进YOLOv8n的火灾检测算法,该算法旨在提高对火灾中火焰和烟雾目标的检测精度。首先,对YOLOv8主干网络中的C2f模块进行重构,设计可变形自注意力残差块(CSP-DAM),以提高主干网络对多尺度、多形变信息的捕获能力。其次,在Neck部分设计小目标检测层,并融合CSP-GSC(CSP-GSConv),通过增加下采样时输出的特征通道数,在降低参数量的同时增强通道间的特征融合,提高小尺度目标的检测精度。最后,融合带有辅助边界框回归机制的Auxiliary-CIOU作为损失函数,以更有效的梯度流动加速收敛过程。在M4SFWD数据集实验中,改进后算法的mAP@0.5达到88.7%,比基线提高1.6%。与其他算法相比,本算法具有更少的参数量、更低的漏检率和更高的检测精度,能够有效提高无人机航拍视角下的火灾巡检效率。In the context of global climate change,forest fires are an increasing threat to the environment and public safety.To address the misdetection and omission of multi-scale and deformable targets in long-distance UAV fire warning systems,an improved fire detection algorithm based on YOLOv8n is proposed,aiming at enhancing the detection accuracy of flame and smoke targets.Firstly,the C2f module in the YOLOv8 backbone network is reconfigured and the deformable self-attentive residual block(CSP-DAM)is designed to enhance the backbone network's ability to capture multi-scale and deformable information.Secondly,a small target detection layer is designed in the Neck section and fused with CSP-GSC(CSP-GSConv)to improve the detection accuracy of small-scale targets.This is achieved by increasing the number of feature channels output during downsampling and enhancing the feature fusion between channels,while also reducing the number of parameters.Finally,we incorporate Auxiliary-CIOU with an auxiliary bounding box regression mechanism as a loss function to accelerate the convergence process through more efficient gradient flow.In the experiments on the M4SFWD dataset,the improved algorithm achieves an mAP@0.5 of 88.7%,which is 1.6%higher than the baseline.Compared with other algorithms,the proposed algorithm has fewer parameters,a lower leakage rate,and higher detection accuracy,significantly improving the efficiency of fire inspection from a UAV aerial perspective.

关 键 词：YOLOv8n 自注意力机制 小目标检测 烟火检测 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]