没有合适的资源?快使用搜索试试~ 我知道了~
温馨提示
试读
18页
Purpose一Fire smoke detection in petrochemical plant can prevent fire and ensure production safety and life safety. The purpose of this paper is to solve the problem of missed detection and false detection in flame smoke detection under complex factory background. Design/methodology/approach一This paper presents a flame smoke detection algorithm based on YOLOvS. The targetregression loss function (CIoLn is used to improve the missed detection and false detection in target detection and improve the
资源推荐
资源详情
资源评论
Flame smoke detection
algorithm based on YOLOv5
in petrochemical plant
Yueting Yang
Guangdong University of Petrochemical Technology, Maoming, China and
Jilin Institute of Chemical Technology, jilin, China, and
Shaolin Hu, Ye Ke and Runguan Zhou
Automation School, Guangdong University of Petrochemical Technology,
Maoming, China
Abstract
Purpose – Fire smoke detection in petrochemical plant can prevent fire and ensure production safety and life
safety. The purpose of this paper is to solve the problem of missed detection and false detection in flame smoke
detection under complex factory background.
Design/methodology/approach – This paper presents a flame smoke detection algorithm based on
YOLOv5. The target regression loss function (CIoU) is used to improve the missed detection and false detection
in target detection and improve the model detection performance. The improved activation function avoids
gradient disappearance to maintain high real-time performance of the algorithm. Data enhancement
technology is used to enhance the ability of the network to extract features and improve the accuracy of the
model for small target detection.
Findings – Based on the actual situation of flame smoke, the loss function and activation function of YOLOv5
model are improved. Based on the improved YOLOv5 model, a flame smoke detection algorithm with
generalization performance is established. The improved model is compared with SSD and YOLOv4-tiny. The
accuracy of the improved YOLOv5 model can reach 99.5%, which achieves a more accurate detection effect on
flame smoke. The improved network model is superior to the existing methods in running time and accuracy.
Originality/value – Aiming at the actual particularity of flame smoke detection, an improved flame smoke
detection network model based on YOLOv5 is established. The purpose of optimizing the model is achieved by
improving the loss function, and the activation function with stronger nonlinear ability is combined to avoid
over-fitting of the network. This method is helpful to improve the problems of missed detection and false
detection in flame smoke detection and can be further extended to pedestrian target detection and vehicle
running recognition.
Keywords Flame smoke detection, Target recognition, YOLOv5, Image detection, Deep learning
Paper type Research paper
1. Introduction
The production safety of the factory has always been a problem that cannot be ignored. Taking
petrochemical plants as an example, due to the dense production workshops and the existence of
a large number of flammable and explosive dangerous goods, once a fire occurs, it is easy to
induce catastrophic consequences, cause environmental pollution and seriously threaten
production safety and personnel life and property safety. Therefore, timely detection and early
warning control of early fires is a realistic demand for safe production. International and domestic
attention has been paid to the flame smoke detection and alarm technology in the factory area.
Early fire detection is mainly achieved through smoke sensors and temperature sensors. For
example, smoke sensors complete fire prevention by detecting smoke concentration. This method
has a good performance in indoor or some small places. However, in a complex environment, due
to the influence of factors such as airflow environment and thermal barrier effect, coupled with
IJICC
16,3
502
This work was supported by National Natural Science Foundation of China (61973094).
The current issue and full text archive of this journal is available on Emerald Insight at:
https://www.emerald.com/insight/1756-378X.htm
Received 18 November 2022
Revised 8 December 2022
Accepted 2 January 2023
International Journal of Intelligent
Computing and Cybernetics
Vol. 16 No. 3, 2023
pp. 502-519
© Emerald Publishing Limited
1756-378X
DOI 10.1108/IJICC-11-2022-0291
the close detection distance of the sensor and low stability, it is difficult to accurately obtain
on-site real-time signal data information by relyingsolelyonthesensorto detect temperature,
concentration and other indicators. Feng and Yang (2019) designed an automatic fire
detection and alarm and automatic fire extinguishing system, the system to achieve intelligent
detection of fire, automatic alarm and real-time accurate fire extinguishing function.
In recent years, with the continuous development of intelligent image processing technology,
especially the application of various models and methods based on deep learning, v arious
intelligent flame smoke detection methods have been widely studied and applied. Many scholars
at home and abroad have applied deep learning technology to flame smoke detection. Ryu and
Kwak (2021) used HSV color conversion and Harris angle detection in the image preprocessing
step and used convolutional neural network to determine the fire area, reducing the missed
detection rate. Jiang and Bai (2019) proposed a flame detection algorithm based on RetinaNet
model, which realized real-time end-to-end flame recognition and positioning. For forest fire
smoke detection, Rager Julia et al. (2021) proposed to increase STN (spatial transformer network)
and entropy function threshold in convolutional neural network to improve the accuracy of
smoke recognition. In order to recognize smoke, Yu (2010) combined the texture features of smoke
and deep learning. Firstly, the author used gray level co-occurrence matrix to recognize the
texture features of smoke and then classified and recognized the smoke blocks by BP neural
network. Through experiments, it is found that the algorithm has a better effect on smoke
recognition. Gao et al. (2018) used the CNN model to identify the smoke image and then combined
the dynamic texture features of the smoke to determine whether it was smoke. This method
reduces the influence of external illumination and smoke concentration changes on model
recognition. Wang (2020a) proposed a smoke detection algorithm using Faster R-CNN. Firstly, the
smoke is extracted by motion detection of the smoke, and then the Faster R-CNN network is used
to extract and identify the smoke image features. Deep learning method is superior to the
traditional artificial feature extraction (Maditham et al., 2022). CNN model has a strong ability to
extract foreground and background features (Huang, 2021; Rao Kota and Devi Munisamy, 2021),
can be used to extract more abstract and deeper features in flame.
Although the research on target detection based on the above methods has made a major
breakthrough, in the actual target detection, the entire image needs to be convoluted,
requiring a larger field of view to meet the algorithm brief and fast. YOLO is a target detector
that uses features learned by deep convolutional neural networks to detect objects. In recent
years, a large number of studies have used it for intelligent detection of different types of
images ( Zhang et al., 2022a). Kris (2020) performed online image enhancement on the self-built
training set by improving the backbone network of YOLOv3 and used the local feature idea to
classify the flame pre-selected area images to achieve smoke detection and flame detection
simultaneously. Ren et al. (2019) used the improved YOLOv3 network to realize the detection
and recognition of fire. The algorithm improves the recognition accuracy and detection speed
of small target smoke by improving the prediction frame size of K-means clustering
algorithm in YOLOv3. Cao et al. (2021) proposed a YOLOv4 accuracy improvement strategy
based on multi-scale feature maps and made some improvements by improving the feature
extraction network to detect small objects.
Xue et al. (2022) proposed an improved model
based on YOLOv5s to improve the accuracy of small target forest fire detection by using
transfer learning method for the small target size in long-distance forest fire images, which is
difficult to capture effective information. However, the model structure is complex and the
flame detection accuracy is not enough (Sukumaran and Brindha, 2020).
In view of the excellent performance of the YOLOv5 algorithm model with small size and fast
detection speed, and due to the complexity of treating the flame smoke detection problem in the
plant area (Zhang et al., 2022b), this article establishes a plant flame smoke detection algorithm on
the basis of specially improved YOLOv5, which is used to solve the problem of missed detection
and false detection of flame smoke. In this algorithm, the original GIoU _ Loss is replaced by CIoU
Smoke
detection
algorithm
503
_ Loss as the loss function of the bounding box, and the SiLU activation function is utilized to
avoid over-fitting of the network. Finally, the effectiveness and accessibility of the proposed
algorithm are verified by comparing with SSD, YOLOv4-tiny and YOLOv7 algorithms.
Specifically, the main research work of this article consists of four sections: Section 1
briefly introduces the network structure of YOLOv5 and elaborates the optimization process
of object detection algorithm based on YOLOv5; section 2, through the independent
construction of two types of data sets, the hyper parameter settings for model preprocessing;
in section 3, three evaluation indexes are selected to evaluate the ability of the network model,
and the algorithm is compared with SSD and YOLOv4-tiny algorithm in the same data set
and experimental environment. Section 4 summarizes the research content of this article and
analyzes the shortcomings of the research as the next stage of the research task. It is verified
that the algorithm in this article takes into account both accuracy and speed in real-time
monitoring of flame smoke, indicating that it has good practicability.
2. Method
2.1 YOLOv5 network model and algorithm improvement
As a new target detection method, YOLO has the characteristics of fast target detection and
high accuracy (Yu, 2022). Therefore, the YOLOv5 model with excellent accuracy has been
greatly popularized in object detection and recognition. The model is improved on the basis of
YOLOv3 and incorporates some new technologies in the field of target detection in recent
years to improve the detection performance of the model (Hua-wei et al., 2022). YOLOv5 is
implemented using the deep learning framework PyTorch, using additional data sets to
enhance training to achieve efficient and accurate recognition of targets. Compared with the
Darknet framework used by YOLOv1 ∼ YOLOv4, the software and hardware support is
relatively complete and supports deployment on a variety of devices. Therefore, YOLOv5 is
selected as the experimental object to detect flame smoke in the plant.
2.1.1 YOLOv5 network model structure. The structure of YOLOv5 can be divided into four
parts: input, backbone, Backbone, Neck and Prediction (Niu et al., 2022). CSP, Focus and SPP
structures are used in the backbone network. The CSP1 _ X and CSP2 _ X structur es in CSP are
applied to Backbone and Neck respectively. CSP solves the network optimization problem of
repeated gradient information in the backbone of other large-scale convolutional neural network
frameworks, reduces the model size and improves the inference speed and accuracy (Fahad et al.,
2022). Focus is a slicing operation on the feature map, which integrates width and height
information into multiple dimensions to improve reasoning speed. The SPP (Spatial Pyramid
Pooling) module performs a maximum pooling of 5 3 5, 9 3 9, 13 3 13 on the image to extract
features from multiple aspects and then concat slices the width and height of the fused image
(Dou et al., 2021). These structures effectively avoid the problems of image distortion caused by
image region clipping and scaling operations and also solve the problem of repeated feature
extraction of images by convolutional neural networks. The Neck layer uses the FPN þ PAN
structure to obtain high-level semantic feature maps and enhance the positioning information and
transmits the image features to the prediction layer to enhance the recognition ability of the model
(Wang, 2022). NMS performs non-maximal suppression on occluded overlapping targets to
obtain the optimal target frame (see Figure 1).
The YOLOv5 adjusted by parameters can be divided into f our models of d ifferent siz es,
namely v5s, v5m, v5l and v5x (Yan-hua et al., 2022). The ov erall stru cture of t he four model s
is the same, and the difference mainly lies in the depth and width of the model, that is, th e
number of model layers and the number of convolutional layers are different, and the
weight file sizes are 14, 42, 93 and 170 MB, respectively. As the network model continues to
deepen, the model detection accuracy continue s to improve, and it has stronger feature
extraction and feature fusion capabilities. As the model deepens , the volume is also
IJICC
16,3
504
gradually i ncreasi ng, resulting in a continuous decline in detection speed and a significant
increase in resource usage. After comparing the performance of each version of YOLOv5
model, this article selects YOLOv5l as the basic mode l with suitable model size, fast speed
and h igh accuracy.
2.2 YOLOv5 algorithm improvement
2.2.1 Improvement of loss function. The loss function is an important indicator for evaluating
regression and classification problems. It is crucial for the back propagation time in deep
learning networks to estimate errors. Therefore, this section improves GIoU _ Loss (Yu et al.,
2021) by introducing a better theoretical CIoU _ Loss (Gu et al., 2022) loss function.
(1) Definition of GIoU loss
In the target detection, it is necessary to compare the detection effect between the detection
box and the real box. The GIoU _ Loss used in the general network solves the problem of the
ratio that cannot be optimized due to the overlap of different target boxes on the basis of IoU
(Yu et al., 2021). The calculation process of GIoU is shown in equation (1):
GIoU ¼ IoU
j
A
c
U
j
j
A
c
j
(1)
IoU ¼
jA \ Bj
jA ∪ Bj
(2)
L
GIoU
¼ 1 GIoU (3)
In equation (2), A and B are the prediction box and the real box respectively, and IoU is the
traditional intersection-over-union ratio, that is, the ratio of the intersection area of the
prediction box and the real box to the union area; A
c
is the area of the minimum bounding
rectangle of the prediction box and the real box; U is the area of the prediction box and the real
box; In equation (3) , L
GIOU
is the loss of GIoU.
Figure 1.
YOLOv5 network
structure
Smoke
detection
algorithm
505
剩余17页未读,继续阅读
资源评论
羁旅少年
- 粉丝: 4719
- 资源: 4
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
最新资源
- 2c60fbb3dt9ad50ed8864298eea1484b.MP4
- 基于yolov8+dlib实现视觉识别的安全驾驶监测系统部署到jetson NX平台源码+模型.zip
- Qt框架+OpenCV+动态爱心+编程教学+520
- 基于opencv+yolov8实现目标追踪及驻留时长统计源码.zip
- 水稻病害基于Yolov8算法优化目标检测识别与AI辅助决策python源码+模型+使用说明.zip
- 海尔618算价表_七海5.20_16.00xlsx(1)(2).xlsx
- WebCrawler.scr
- 【计算机专业毕业设计】大学生就业信息管理系统设计源码.zip
- YOLO 数据集:8种路面缺陷病害检测【包含划分好的数据集、类别class文件、数据可视化脚本】
- JAVA实现Modbus RTU或Modbus TCPIP案例.zip
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功