基于YOLOv8的火灾烟雾检测算法研究

Research on Fire Smoke Detection Algorithms Based on YOLOv8

早期火灾预警为人类的生命财产安全提供有效保障,为了提高算法在复杂场景中对小火焰和烟雾的检测性能,对v8版本的YOLO算法进行改进,设计了一种轻量型的Fire-YOLOv8火灾检测网络。该网络在YOLOv8的基础上增加一个更小的目标检测层,并使用Focus层对输入图像进行切片操作,解决微小火焰检测的难题。在网络优化中,特征提取选用轻量级的BottleneckCSP模块,使用样本数据集进行迁移学习,更新网络参数,能够有效区分火焰、烟雾等干扰信息。实验结果表明:预训练生成的Fire-YOLOv8n火灾检测模型的精确率达到97.1%,mAP@0.5达到95.7%,检测速度达到192 FPS,模型大小仅为8.4 MB,综合性能得到明显提升,可以很好地满足嵌入式设备实时检测的应用需求。

Early fire warning provides effective safeguards for the safety of human life and property.In order to improve the algorithm's detection performance for small flames and smoke in complex scenarios,the v8 version of the YOLO algorithm has been improved to design a lightweight Fire-YOLOv8 fire detection network.The network adds a smaller target detection layer on the basis of YOLOv8 and uses the Focus layer to cut out the input image to solve the puzzle of small flame detection.In network optimization,feature extraction selects lightweight BottleneckCSP modules,uses sample datasets to migrate learning,updates network parameters,and effectively distinguishes interference information such as flame and smoke.The experimental results showed that the pre-trained Fire-YOLOv8n fire detection model accuracy reached 97.1%,mAP@0.5 at 95.7%,the detection speed reached 192 FPS,the model size was only 8.4 MB,and the overall performance was significantly improved,which can well meet the application requirements for real-time detection of embedded devices.