| 摘 要: 高校实验室安全事故的统计表明,火灾是发生频次最高的事故类型。针对传统视觉检测准确率低的缺陷和实验室监控的轻量化的需求,提出基于改进YOLOv5的实验室火灾检测算法YOLOv5s-EBW。首先,在骨干网络SPPF层前引入ECA注意力机制,增强模型对火焰关键特征的提取以抑制背景干扰。其次,将颈部网络的结构替换为加权双向特征金字塔网络BiFPN,优化火焰多尺度特征的融合以提升对小火焰的检测能力。最后,将损失函数调整为WIoU,通过动态加权机制提升火焰的定位精度与训练稳定性。在自建的实验室火焰数据集上进行实验分析,YOLOv5s-EBW模型的mAP_0.5达到93.8%,较原始YOLOv5s提升了4.4%,而参数量仅增加了2.1%,在保持原始模型轻量化特点的同时提升了火灾检测精度,满足实验室火灾检测要求。 |
| 关键词: 火灾检测 YOLOv5 注意力机制 双向特征金字塔网络 损失函数 |
|
中图分类号:
文献标识码:
|
| 基金项目: 广东省高等教育教学改革项目“基于云服务器的学生、设备、实验室三位一体智慧管理系统”；2024年校级质量工程项目“电子信息工程专业课程教研室”（ZLGC202423） |
|
| Research on a Laboratory Fire Detection Algorithm Based on Improved YOLOv5CHEN Ruikun1 LI Renxun1 LAI Changchen1 XU Qinghua1 ZHONG Dongyu2 SUN Daozong1 |
|
Chen Ruikun, Li Renxun, Lai Changchen, Xu Qinghua, Zhong Dongyu, Sun Daozong
|
South China Agricultural University
|
| Abstract: Statistics on safety accidents in university laboratories indicate that fires are the most frequent type of hazard. To address the low accuracy of traditional vision-based fire detection approaches and the demand for lightweight deep learning models in laboratory monitoring, this paper proposes a laboratory fire detection algorithm named YOLOv5s-EBW based on an improved YOLOv5 framework. In the backbone network, the ECA attention mechanism is introduced before the SPPF layer to enhance the extraction of key flame features and reduce background interference. The original neck network is replaced by the weighted bidirectional feature pyramid network (BiFPN) to optimize multi-scale feature fusion and improve the detection capability for small flame targets. Furthermore, the loss function is modified to WIoU, which boosts flame localization accuracy and training stability through a dynamic weighting mechanism. Experimental results on a self-built laboratory flame dataset show that the mAP_0.5 of YOLOv5s-EBW reaches 93.8%, 4.4% higher than the original YOLOv5s, with only a 2.1% increase in parameters. The model retains the lightweight advantage of the baseline network while significantly improving detection precision, fully satisfying the fire detection requirements of university laboratory scenarios. |
| Keywords: Fire detection YOLOv5 Attention mechanism Bi-directional Feature Pyramid Network Loss function |
