基于改进YOLO-V5 算法的烟火检测方法
Firework detection method based on improved YOLO-V5 algorithm
张明振 1段江忠 1梁肇伟 2郭俊杰 1柴大山3
作者信息
- 1. 深圳市城市公共安全技术研究院有限公司,广东 深圳 518038
- 2. 深圳技术大学 大数据与互联网学院,广东 深圳 518118
- 3. 中国铁塔股份有限公司 深圳市分公司,广东 深圳 518000
- 折叠
摘要
为减少自然环境中云、水雾、沙尘、灯光、日出、日落等干扰因素对烟雾、火焰目标检测准确性的影响,提出一种基于改进YOLO-V5 算法的烟火检测算法.采用现场采集和网络爬取的方法获取烟雾、火焰目标图像和干扰类图像数据集,均衡学习训练样本,提高模型泛化能力;使用加权双向特征金字塔网络(BiFPN)替换原有的特征金字塔网络(FPN)+路径聚合网络(PAN)结构,对目标进行多尺度特征融合,加强模型特征融合能力;同时,运用距离交并比(DIoU)非极大值抑制(NMS)替代原有的NMS,加快检测框损失函数收敛速度,加强模型推理能力.结果表明:改进后的算法准确率为79.2%,召回率为68.6%,平均精度均值(mAP)为74.2%,误报率(FPR)为12.8%;相比于原YOLO-V5 算法,改进后的算法准确率、召回率、mAP 分别提高 1.9%、0.9%、2.7%,检测识别FPR降低3.7%.
Abstract
To reduce the influences of background interference factors in natural environments such as clouds,mist,dust,lights,sunrise,and sunset on the smoke and flame target detection accuracy,a smoke and fire detection algorithm based on an improved YOLO-V5 algorithm was proposed.Smoke,flame target images,and interference image data sets were obtained from the on-site collection and web crawling approaches to solve sample imbalance and improve model generalization ability.A bidirectional feature pyramid network(BiFPN)was used to replace the original feature pyramid network(FPN)+path aggregation network(PAN)structure,and then multi-scale feature fusion on the target was performed to strengthen the model feature fusion ability.At the same time,distance intersection-over-union(DIoU)non-maximum suppression(NMS)is used to replace the original non-maximum suppression(NMS)to speed up the convergence of the detection box loss function and enhance the model reasoning ability.The results showed that the improved algorithm's accuracy,recall rate,mean average precision(mAP)and FPR were 79.2%,68.6%,74.2%,and 12.8%,respectively.Compared with the original YOLO-V5 algorithm,the proposed algorithm improved accuracy rate,recall rate,and mAP by 1.9%,0.9%,and 2.7%,respectively.Furthermore,the FPR was decreased by 3.7%.
关键词
YOLO-V5/算法/烟雾/火焰/目标检测/误报率(FPR)Key words
YOLO-V5 algorithm/smoke/fire/target detection/false positive rate(FPR)引用本文复制引用
出版年
2024
NETL
NSTL
万方数据