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。
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.
作者
康斓
蒋晓刚
苑志江
KANG Lan;JIANG Xiaogang;YUAN Zhijiang(Dalian Naval Academy,Dalian 116018,China)
出处
《通信与信息技术》
2024年第2期21-26,共6页
Communication & Information Technology
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
Fire smoke detection
Object detection
Attention mechanism
Feature fusion
Ship safety
