为解决森林防火工作中监测预警不及时、不准确的问题，基于YOLOv8模型提出了一种面向复杂场景的监测技术，实现了森林火情的快速准确监测。在Backbone部分引入动态蛇形卷积(Dynamic Snake Convolution，DySConv)模块和全局注意力机制(Global Attention Mechanism，GAM)模块，显著增强了模型的特征提取能力，提升了烟火特征的表现力。采用WIoU(Wise Intersection over Union)损失函数，增强网络对普通质量锚框的关注度，从而提供更准确的目标检测评估。通过将DyHead模块集成到检测头中，增强了检测头的尺寸、空间和任务感知能力，优化了整体性能。为了对模型性能进行严格评估，设计了消融实验和主流模型对比实验，结果表明，提出的方法是有效的，该方法的权重大小为 14。4 MB，平均精度均值(mean Average Precision，mAP)@。5:。95 达到 80。3％。相较于 YOLOv8，mAP@5:。95 提高了 8。7％。该模型可为远程火情监控与预警提供技术支持，在森林防火工作中具备实用价值和现实意义。
Research on Forest Fire Monitoring Technology for Complex Scenarios
To solve the problem of untimely and inaccurate monitoring and warning in forest fire prevention,this paper proposes a complex scene-oriented monitoring technique based on the YOLOv8 model,which achieves fast and accurate monitoring of forest fire.In the backbone section,the method incorporates a Dynamic Snake Convolution(DySConv)module and a Global Attention Mechanism(GAM)module,significantly enhancing feature extraction capability and improving the representation of firework features.Additionally,by utilizing Wise Intersection over Union(WIoU)loss function,attention to anchor boxes with moderate quality is enhanced,providing a more precise assessment of the target detection.Integrating the DyHead module into detection head enhances the scale,spatial,and task awareness of the detection head,optimizing overall performance.To rigorously evaluate the performance,this paper designed ablation experiments and comparative experiments with mainstream models.Experiment results indicate that the proposed model is effec-tive,with a model weight size of 14.4 MB,and the mean Average Precision(mAP)@.5:.95metric reaching 80.3％.Compared to YOLOv8,the mAP@.5:.95 value is increased by 8.7％.The model can provide technical support for remote forest fire monitoring and early warning,which has practical value and immediate significance in forest fire prevention work.
国家科技期刊平台
NSTL
万方数据
