融合多特征的火焰和烟雾深度视觉检测模型研究
A Deep Flame and Smoke Visual Detection Modol Research By Fusing Multi-Scale Features
周仿荣 1马仪 1文刚 1王雷光 2刘太文 1黄灏3
作者信息
- 1. 云南电网有限责任公司电力科学研究院,云南 昆明 650217
- 2. 西南林业大学园林园艺学院,云南 昆明 650224
- 3. 上海师范大学信息与机电工程学院,上海 201418
- 折叠
摘要
为了同时准确地检测火焰和烟雾目标,开展了多特征增强融合的视觉火灾探测网络模型的研究.在多尺度特征提取骨干网络的基础上,提出高层语义空间信息增强模块、多尺度特征深度融合模块、空间距离信息注意力模块、分类定位模块.这些模块分别聚焦烟雾和火焰的空间纹理、多尺度特征、视觉注意力以及小火焰和烟雾目标检测等问题.本文参与建构一个新的火焰和烟雾图像检测数据集,方便火灾检测模型的训练和测试.实验结果表明,本文方法取得的检测指标超过对比算法.
Abstract
To accurately detecting flame and smoke simultaneously,we fuse and enhance features to propose a deep visual fire detection network.On the basis of a multi-scale feature backbone network,we design a high-level semantic spatial information enhancement module,a multi-scale feature depth fusion module,a spatial distance information attention module,and a classification and positioning module.These modules mainly focus on extracting spatial textures and multi-scale features of flame and smoke,and detecting small objects of flame and smoke.We participated in constructing a new flame and smoke detection dataset for training and testing visual fire detection models.Experimental results show that our method achieves better results than the compared methods.
关键词
火灾探测/火焰检测/烟雾检测/特征增强/目标检测Key words
Fire Detection/Flame Detection/Smoke Detection/Feature Enhancement/Object Detection引用本文复制引用
基金项目
云南省重大科技专项(202202AD080010)
出版年
2024
NETL
NSTL
万方数据