基于改进YOLOv5的轻量化早期舰船火灾烟雾检测算法
Early ship fire and smoke detection based on improved YOLOv5
康斓 1蒋晓刚 1苑志江1
作者信息
- 1. 海军大连舰艇学院,辽宁大连 116018
- 折叠
摘要
为改进当前目标检测技术对舰船环境下早期火灾烟雾小目标的检测精度不足、速度较慢与特征表达能力不充分的现状,对改进YOLOv5的轻量化早期舰船火灾烟雾检测方法进行了研究.首先,对YOLOv5网络结构进行分析;其次,从主干网络MobileNetv3架构轻量化、检测颈Bi-FPN特征融合金字塔改进、优化Alpha-IoU损失函数与改进Soft-NMS四方面改进YOLOv5算法;最后,构建消融与对比实验验证改进性能.改进模型训练时间减少34.1%,mAP较原模型提升了6.6%.其对早期舰船火灾烟雾检测效果优异,对舰船火灾的早期预警具有指导意义.
Abstract
In order to improve the current target detection technology's insufficient accuracy,slow speed and insufficient feature expression ability for early fire smoke detection of small targets in ship environment,the lightweight early ship fire smoke detection method with improved YOLOv5 was studied.Firstly,the network structure of YOLOv5 is analyzed.Secondly,the YOLOv5 algorithm is improved from four aspects:lightweight architecture of backbone network MobileNetv3,improvement of detection neck Bi-FPN feature fusion pyramid,optimization of Alpha-IoU loss function and improvement of Soft-NMS.Finally,ablation and comparison experiments were constructed to verify the improved performance.The training time of the improved model is reduced by 34.1%,and the mAP is im-proved by 6.6%compared with the original model.Its effect on early ship fire smoke detection is excellent,and it has guiding signifi-cance for the early warning of ship fire.
关键词
火灾烟雾检测/目标检测/注意力机制/特征融合/舰船安全Key words
Fire smoke detection/Object detection/Attention mechanism/Feature fusion/Ship safety引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据