| 摘 要: 针对复杂环境中,火焰检测存在特征提取不足和边缘模糊目标检测性能欠缺问题,提出一种融合挤压激励(Squeeze-and-Excitation,SE)注意力机制的YOLOv7火焰目标检测算法。该算法以YOLOv7为基础框架,基于公开火焰数据集,对不同位置点插入SE注意力机制的网络模型进行研究,进而构建融合多点注意力机制的YOLOv7_Attention网络模型,以充分提取火焰的有效特征,抑制冗余特征。实验结果表明,融合SE注意力机制的YOLOv7_Attention网络模型与原始YOLOv7模型相比,其mAP提升了1.64百分点,边缘模糊火焰目标检测效果显著。 |
| 关键词: 火焰图像;目标检测;注意力机制;YOLOv7 |
|
中图分类号: TP391
文献标识码: A
|
| 基金项目: 山西科技学院校内科研基金项目(XKY002) |
|
| Research on YOLOv7 Fire Target Detection Algorithm Based on Multi-location Attention Mechanism |
|
ZHANG Dongmei1, SONG Zitao2, FAN Haoxin1
|
(1.School of Big Data and Computer Science, Shanxi Institute of Science and Technology, Jincheng 048000, China;
2.School of Opto-electromechanical Engineering, Shanxi Institute of Science and Technology, Jincheng 048000, China)
zhangdongmei@sxist.edu.cn; songzitao@sxist.edu.cn; fanhaoxin@sxist.edu.cn
|
| Abstract: Aiming at the issues of insufficient feature extraction and lack of detection performance for fuzzy edge targets in flame detection in complex environments, this paper proposes a YOLOv7 flame target detection algorithm integrated with a Squeeze-and-Excitation (SE) attention mechanism. Based on the YOLOv7 framework and utilizing a publicly available flame dataset, the study investigates a network model with SE attention mechanisms inserted at different locations, thereby constructing the YOLOv7_Attention network model integrated with a multi-point attention mechanism to effectively extract valid features of flames and suppress redundant features. Experimental results show that the mAP of the YOLOv7_Attention network model integrated with SE attention mechanism improves by 1.64 percentage points compared to the original YOLOv7 model, significantly enhancing the detection performance of fuzzy edge flame targets. |
| Keywords: fire image; target detection; attention mechanism; YOLOv7 |
