TITLE:
Improved High Speed Flame Detection Method Based on YOLOv7
AUTHORS:
Hongwen Du, Wenzhong Zhu, Ke Peng, Weifu Li
KEYWORDS:
Light Weight, Detection of Flame, YOLOv7-CN-B, YOLOv7, ConvNeXt
JOURNAL NAME:
Open Journal of Applied Sciences,
Vol.12 No.12,
December
14,
2022
ABSTRACT: In
order to solve the problems of the traditional flame detection method, such as
low detection accuracy, slow detection speed and lack of real-time detection
ability. An improved high speed flame detection method based on YOLOv7 is
proposed. Based on YOLOv7 and combined with ConvNeXtBlock, CN-B network module
was constructed, and YOLOv7-CN-B flame detection method was proposed. Compared
with the YOLOv7 method, this flame detection method is lighter and has stronger flame feature
extraction ability. 2059 open flame data sets labeled with single flame
categories were used to avoid the enhancement effect brought by high-quality
data sets, so that the comparative experimental effect completely depended on
the performance of the flame detection method itself. The results show that the
accuracy of YOLOv7-CN-B method is improved by 5% and
mAP is improved by 2.1% compared with YOLOv7 method. The detection speed
reached 149.25 FPS,
and the single detection speed reached 11.9 ms. The experimental results show
that the YOLOv7-CN-B method has better performance than the mainstream
algorithm.