基于YOLOv7的红外阴燃火探测算法改进研究  

Improvement of Infrared Smoldering Fire Detection Algorithm Based on YOLOv7

作　　者：汤伟[1] 张文迪 袁航 解聪 任家辉 Tang Wei;Zhang Wendi;Yuan Hang;Xie Cong;Ren Jiahui(School of Electrical and Control Engineering,Shaanxi University of Science and Technology,Xi’an 710021,China)

机构地区：[1]陕西科技大学电气与控制工程学院,西安710021

出　　处：《燃烧科学与技术》2024年第5期532-538,共7页Journal of Combustion Science and Technology

摘　　要：目前,基于机器视觉的火灾检测算法中数据集类型不充分、数据集在时间维度覆盖不全面,致使此类算法难以实现火灾的早期预警,文中提出了基于改进YOLOv7的红外阴燃探测方法.该算法利用EfficientFormerV2模型替换原模型的骨干网络CSPDarknet53,从而增强了模型低延迟、低参数量、易部署的能力;同时,在预测网络中,采用CARAFE轻量化上采样模块代替原模型中的上采样模块,扩大了模型对特征的感受野,改善了阴燃特征的表示能力;此外,还引入了新的NWD度量来提升模型边界框预测能力.结果表明,在自建阴燃数据集上,该算法的平均精度达到92.9%,对阴燃检测的平均精度达到99.6%,比YOLOv7的精度提升了14.4%,较基于手工提取特征的卷积神经网络算法提升了4.6%.研究成果将为阴燃火早期预警提供新思路.Insufficient dataset types and incomplete coverage of datasets in time dimension are not considered in the current machine vision-based fire detection algorithms.To address the difficulty in achieving the early warning of fires,the paper proposes a method based on improved YOLOv7 for detecting infrared smoldering fires.The algorithm uses the EfficientFormerV2 model to replace the CSPDarknet53 backbone network,thus improving the model’s capability in terms of low latency,low parameter count,and easy deployment.Meanwhile,the prediction network uses the CARAFE lightweight up sampling module to replace the UP module in the original model,expanding the model’s receptive field of the features and improving the model’s representation capabilities of smoldering fire features.The approach achieves 92.9%mAP on the self-constructed smoldering fire dataset and 99.6%mAP for smoldering fire,which is 14.4%higher than the mAP of YOLOv7 and 4.6%higher than the accuracy of the convolutional neural network algorithm based on hand-extracted features.The findings of the study will bring new ideas for the early warning of smoldering fires.

关 键 词：阴燃火 火灾检测 YOLOv7 红外探测 

分 类 号：TN215[电子电信—物理电子学]