基于改进YOLOv7的火灾火焰检测模型

打开文本图片集

关键词：YOLOv7；火灾检测；Wise-IoU损失函数；注意力机制；PConv模块 中图分类号：TP391 文献标志码：A 文章编号：1671-6841（2025)05-0001-08 DOI:10.13705/j. issn.1671-6841.2024041

A Flame Detection Model Based on Improved YOLOv7

LIU Chengming，WU Fan，LI Xuexiang ( School of Network Technology Security， Zhengzhou University， Zhengzhou 45Oo02，China)

Abstract:The issue of inadequate effectivenessof existing object detection models for detecting flame was addressed.A novel fire detection model was proposed by optimizing and enhancing the network structure of the YOLOv7 algorithm. Several innovative improvements were introduced in three key aspects.Firstly, the YOLOv7 was augmented with a small object detection layer and SE attntion，coordinate attention, and Biformer modules were incorporated to enhance the extraction of smallobject features. Secondly，the CoordConv and PConv modules were integrated to replace the standard convolution blocks in the network, resulting in reduced computational complexity during training and deployment，and improved network detection speed.Lastly，the issue of inconsistent quality of annotated bounding boxes in the experimental dataset was addressed by replacing the CIoU loss function with the Wise-IoU loss function. Experimental results conducted on the KMU Fire and Smoke database demonstrated that the improved model achieved a 2.5 percentage points increase in average precision and a 1.7 percentage points increase in recall. Additionally，the frame rate reached 79.4 frames per second. This dual improvement in detection performance and speed surpassed that of traditional object detection algorithms，making the model more effective in detecting fires.

Key words: YOLOv7； fire detection； Wise-IoU loss function; atention mechanism； PConv module

0 引言

灭灾作为一种对人们生命财产安全有着巨大威胁的灾害，其爆发往往是突然的、迅猛的。（剩余17191字）