基于CG-yolo的烟火检测

Detection of Fireworks Based on CG-yolo

针对山火烟雾的检测存在由于监控范围广、发生频率不固定等造成的高成本问题,在边缘计算思维的启发下,提出了一个基于YOLOv5改进的适用于前端布设的轻量级识别网络。该方法针对YOLOv5模型过大的缺陷,通过修改网络结构,将融合了通道注意力机制CoordAttention的Ghostbottleneck模块与YOLOv5结合,提出一种改进型卷积神经网络CG-yolo识别网络。实验结果表明,CG-yolo相对于YOLOv5s算法速度提高了9.5%,查全率提升了1.8%,查准率仅损失1.7%,部署在NVIDIA的Jetson Nano边缘计算平台上时运行速度可以达到13fps,更好地满足了隐患监测的工程实际需求。

Aiming at the high cost problem of mountain fire smoke detection caused by wide monitoring range and irregular oc⁃currence frequency,inspired by the edge computing thinking,this paper proposes a lightweight recognition network based on YO⁃LOv5,which is suitable for front-end layout.In order to solve the problem of too large model of YOLOv5,this paper proposes an im⁃proved convolutional neural network CG-yolo recognition network by modifying the network structure and combining the ghost bottle⁃neck module of coordattention with YOLOv5.The experimental results show that CG-yolo can improve the speed by 9.5%,recall by 1.8%and precision by 1.7%compared with YOLOv5s algorithm.When deployed on NVIDIA's Jetson nano edge computing plat⁃form,the running speed can reach 13fps,which can better meet the practical needs of hidden danger monitoring.