Fire And Smoke Detection Using Deep
Learning
ICNWC_101
1stDivya B 2ndKavitha S 3rdMuthu Pandeeswari R
Assistant Professor, Dept of Computer Assistant Professor, Dept of Assistant Professor, Dept of
Science and Computer Science and Computer Science and
Engineering, Sri Sai Ram Institute of Engineering(Cyber Security), Sri Sai Engineering, Rajalakshmi
Technology Ram Institute of Engineering College, Thandalam
[email protected] Technology [email protected]
[email protected]
4thSreenidhi R 5thPonvaishnavi MR 6thLoshine S
UG Scholar, Dept of Computer Science UG Scholar, Dept of Computer UG Scholar, Dept of Computer
and Engineering, Science and Engineering, Science and Engineering,
Sri Sai Ram Institute of Technology Sri Sai Ram Institute of Technology Sri Sai Ram Institute of Technology
[email protected] [email protected] [email protected]
ABSTRACT
• Of late, wildfires a n d c o m m e r c i a l f i r e s l i k e f i r e i n s h o p p i n g c o m p l e x , f i r e
w o r k f a c t o r i e s a n d i n d u s t r i e s , continue to cause extensive destruction throughout the
world, frequently causing human fatalities. The solution is divided into two systems, where the
first system focuses on the wildfire and the second system focuses on the commercial fires.
• Identifying the fire and smoke correctly plays an important role. Gradient-weighted Class
Activation Mapping (Grad-CAM) is used , to ensure that detections corresponded to the smoke
regions in an image.
• The detection algorithms tested were residual neural networks (ResNet), CNN(Convolutional
neural network) and VGG16 (Visual Geometry Group) algorithms of various sizes because
these algorithms have given good results in the past in multiple domains .The training was done
using transfer learning.
• The Gradient-weighted Class Activation Mapping (GradCAM), a visual explanation algorithm, is
used to test images originating alarms, which allows confirming that the neural network’s
‘‘attention’’ (region of interest) is on the fire location and not somewhere else. The GradCAM
outcome is the result of passing this sum by a ReLU activation function. This ReLU is used to
show only the pixels that have a positive influence on the desired outcome.
• The saliency map is also used in finding out the difference between smoke and the fog in case
of wild fire.
ICNWC2024 2
ABSTRACT
• In case of the building fires like shopping complex, fire work shops,
factories, industries would also require a better solution because the
detection is easier but the rescue process is more difficult in such places.
• The same procedure is followed for detecting and finding origin of the fire as
the wild fire model. The additional feature is added to the model for rescue
purpose. The fire control and the evacuation becomes more complex and
hence the solution to such problem is not yet brought into the consideration.
• Even after detecting the fire, most of the people die due to the poor rescue
process. This is because, the rescue team cannot find the people inside the
building. Our solution takes that into the consideration. Our system uses the
camera which could alert the people inside and also helps in the rescue
process.The camera takes the photo and video of the area if the fire or smoke
is detected.
• Then the De-hazing algorithm is used to clarify the smoky images to
identify the core details inside it. This is passed to the rescue team with the
video of people inside which is given the at most priority for extinguish
purpose. The separate mobile application is made for the rescue team.
ICNWC2024 3
INTRODUCTION
• Despite all scientific advances, wildfires and commercial fires like fire in shopping malls, fire work factories and industries, continue to cause
massive destruction, frequently causing human fatalities.
• It is common to see gigantic wildfires in Australia, the United States of America, Spain, or Portugal during summer. Various methods have been
developed to detect wildfires as soon as they start. It is fundamental to attack the flames before they spread too much to extinguish them with
relative ease. Hence identifying the fire and smoke correctly plays an important role. The existing model does not focus on finding the
origin.
• The second problem with the existing model is that it could not find the difference between the fog and the smoke in case of wild fires.
• The third problem in the current model is that it is not compatible with both wildfire and commercial fires. Our solution overwhelms the
problem.
• In case of the building fires like shopping complex, fire work shops, factories, industries would also require a better solution because the detection is
easier but the rescue process is more difficult in such places.
• The same procedure is followed for detecting and finding origin of the fire as the wild fire model. The additional feature is added to the model for
rescue purpose.
• The fire control and the evacuation becomes more complex and hence the solution to such problem is not yet brought into the consideration.
• Even after detecting the fire, most of the people die due to the poor rescue process. This is because, the rescue team cannot find the people inside
the building.
ICNWC2024 4
LITERATURE SURVEY
1) Armando M. Fernandes, Andrei B. Utkin, Paulo Chaves “Automatic Early Detection of Wildfire Smoke With Visible Light Cameras Using
Deep Learning and Visual Explanation”, This system uses the Resnet algorithm and the use of Gradcam and mosaic model to increase
the efficiency. Resnet was not so efficient for real time implementation.
2) Dali Sheng, Jinlian Deng And Jiawei Xiang “Automatic Smoke Detection Based on SLIC-DBSCAN Enhanced Convolutional Neural
Network” In this paper, a hybrid of SLIC-DBSCAN and CNN model is proposed to recognize fire and smoke modes connected to fire
images to generate an automatic smoke detection scheme. SLIC has the strong ability to obtain spectral feature of smoke areas
represented by super-pixel images. Not always work well for the mixture of super-pixel images to distort the recognition precision.
3) Zhijian Yin, Boyang Wan , Feiniu yuan , Xue Xia and Jinting shi “A Deep Normalization and Convolutional Neural Network for Image
Smoke Detection” This paper uses batch normalization to speed up the training process and boost accuracy of smoke detection. Unlike
algorithms based on handcrafted features, our DNCNN can automatically extract features for smoke detection. Experimental results
show that our DNCNN achieves high detection rates and low false alarm rates at the same time.
4) Xiwen chen, Bryce Hopkins , Hao Wang , Leo O’Neill, Fatemeh Afghah and Abolfazi “Wildland Fire Detection and Monitoring Using a
Drone-Collected RGB/IR Image Dataset” A dual-feed prescribed fire imaging dataset collected by unmanned aerial systems in a
ponderosa pine forest using side-by-side visual and thermal camera feeds. In addition, we present and apply our new deep learning-based
analysis methods to the dataset to accurately label and segment, frame by frame, pixels with fire and/or smoke present.
5) Jie Shi,Wei Wang , Yuanqi Gao, Nanpeng Yu “ Optimal Placement and Intelligent Smoke Detection Algorithm for Wildfire-Monitoring
Cameras” This paper proposes a lightweight physics-based video smoke detection framework and an optimal placement strategy for
wildfire camera applications. The proposed smoke detection algorithm extracts useful features of videos and embeds them in LBP-
motion images. This approach enables us to leverage lightweight deep convolutional neural network to accurately perform video smoke
detection with limited computational and storage resources. The actual coverage of a camera can be less than the area of the circle due to
occlusion.
ICNWC2024 5
•Cannot differentiate between the smoke and the fog
LITERATURE SURVEY
6) Gao Xu, Qixing Zhang, Dongcai Liu, Gaohua Lin and Jinjun Wang ”Adversarial Adaptation From Synthesis to Reality in Fast Detector
for Smoke Detection” This paper proposed an end-to-end trainable framework based on fast detector SSD and MSCNN for smoke
detection. To minimize the domain shift between synthetic and real smoke samples for smoke detection, we incorporate the domain
adaptation into the state-of-the-art fast detector (SSD and MSCNN). ZFNet is more rigorous to detect smoke which cause the high
missing detection error but little confusing detection error.
7) Zhe Liu, Xiaoping Yang, Ying Liu and Zhihong Qian “Smoke-Detection Framework for High-Definition Video Using Fused Spatial- and
FrequencyDomain Features” the ViBe algorithm is used to extract the motion regions, the number of regions to be classified is reduced
using morphological processing and filtering of image blocks, and, finally, extracted small image blocks are fed to the SVM for
classification.
8) Feiniu Yuan, Xue Xia, Jinting Shi, Hongdi Li and Gang Li “Non-Linear Dimensionality Reduction and Gaussian Process Based
Classification Method for Smoke Detection” Improve smoke classification accuracy by non-linearly mapping LBP-like features into a
low-dimensional space and modeling classification as a Gaussian Process. We extract LBP-like features, which are one of the best texture
descriptors
9) Yanbei Liu, Wen Qin, Kaihua Liu, Fang Zhang and Zhitao Xiao ”A Dual Convolution Network Using Dark Channel Prior for Image
Smoke Classification” a deep convolutional neural network joint dark channel is proposed to detect smoke images. For the first channel,
the residual network is added on the AlexNet to improve the network and extract generalization features. The dark channel is employed in
the second channel, which inputs the processed dark channel images and extracts the dark channel features.
10)The residual was less efficient and there was false predictions
•The accuracy was lesser and there was more false predictions.
ICNWC2024 •Cannot differentiate between the smoke and the fog 6
PROPOSED METHODOLOGY
● The Gradient-weighted Class Activation Mapping (GradCAM) , a visual explanation algorithm, is used to test
images originating alarms, which allows confirming that the neural network’s ‘‘attention’’ (region of interest) is on
the fire location and not somewhere else.
● This is relevant since it is frequent to find excellent results in the literature that cannot be trusted due to being
obtained with small datasets.
● The GradCAM outcome is the result of passing this sum by a ReLU activation function. This ReLU is used to
show only the pixels that have a positive influence on the desired outcome.
● The purpose is to provide the neural network with information about the surroundings of the smoke plumes
that must be detected. A narrower field of-view can easily lead to misinterpretation of what is visible in the image.
● The main objective of the proposed system is to provide the detection easier with greater accuracy.
● This strategy also has the advantage of providing a rough location of the smoke plume position in the image. The
Saliency approach does not require changing the input image before applying the neural network; only the output
differs.
ICNWC2024 7
PROPOSED METHODOLOGY
● The GradCAM outcome is the result of passing this sum by a ReLU activation function. This ReLU is used to show
only the pixels that have a positive influence on the desired outcome.
● The purpose is to provide the neural network with information about the surroundings of the smoke plumes that must
be detected. A narrower field of-view can easily lead to misinterpretation of what is visible in the image.
● The conventional studies used only the Residual neural network .The accuracy can be increased by using both the
Residual neural network and CNN.
● The VGG16 is another algorithm for deep learning which can be used for increasing the efficiency and accuracy.
● VGG16 is one of the most advanced neural network structures created up to now. Their leap forward was realizing that
scaling in terms of width ,depth.
● Saliency map is also used for visualization. It is an image that highlights the region on which people’s eyes focus
first. The importance of saliency map is to reflect the degree of importance of a pixel to human visual system.
ICNWC2024 8
PROPOSED METHODOLOGY
• The Node mcu is connected with the carbon monoxide sensor to detect the carbon monoxide released from the
smoke. These detectors are made to monitor the amount of carbon monoxide in the air in parts per million, or
ppm, and to alert users if the gas concentration rises to a point where it poses a threat. A few hours of monitoring for
carbon monoxide at 70 parts per million will sound an alert.
• The camera is also connected in the bread board. Python program is written to communicate with the sensors. Using
sensor data to determine when to turn on the camera to take pictures or record videos. For instance, you may take
a picture when a temperature sensor reaches a threshold level or when a motion sensor notices movement.
• A solution that can find and follow smoke patterns more easily using de-hazing algorithm to increase the visibility of
buildings and objects in photos or videos that have been damaged by smoke. By using de-hazing techniques, the
quality of images taken in wildfire-prone areas may be improved, making it simpler to spot and track the spread of smoke
and flames.
ICNWC2024 9
PROPOSED METHODOLOGY
• The application is created for the rescue team in
case of the fire alert in buildings. The application
is designed in such a way that the video from the
camera (In Commercial Buildings) is sent directly
to the rescue team through the application.
• The application shows the people trapped using
the de-hazing algorithm and finally This helps
the rescue team to rescue the people trapped in
the fire.
• The team in-turns safe guard the people
ICNWC2024 10
PROPOSED METHODOLOGY
ICNWC2024 11
PROPOSED METHODOLOGY
ICNWC2024 12
PROPOSED METHODOLOGY
ICNWC2024 13
RESULTS AND DISCUSSION
CNN ResNet VGG 16
Accuracy 96.88% 96.88% 99.22%
ICNWC2024
ACCURACY 14
RESULTS AND DISCUSSION
ICNWC2024
SALIENCY MAP OUTPUT 15
RESULTS AND DISCUSSION
GRADCAM OUTPUT
ICNWC2024 16
RESULTS AND DISCUSSION
DE-HAZING ALGORITHM
ICNWC2024 17
CONCLUSION
● To wrap things up,CNN, ResNet and VGG 16 along with yolo v8 are used to
increase the accuracy.The visualization techniques like GradCAM, and Saliency
map are used to find the origin of the fire and to find out the difference between fog
and smoke in case of wild fire. In commercial buildings, The CNN, ResNet and
VGG 16 ,yolov8 are used to increase the accuracy and visualization techniques
are also used. Additionally De-hazing algorithm is used to clarify the smoky
images and used for rescue team. The rescue team is notified with the alert if the
fire detected. Hence by using these ideas, the efficiency can be increased.
ICNWC2024 18
REFERENCES
[1] Armando M. Fernandes, Andrei B. Utkin, Paulo Chaves “Automatic Early Detection of Wildfire Smoke With Visible Light Cameras Using Deep Learning and Visual
Explanation”, Date of publication:25 January 2022, Doi: 10.1109/ACCESS.2022.3145911.
[2] Dali Sheng, Jinlian Deng And Jiawei Xiang “Automatic Smoke Detection Based on SLIC-DBSCAN Enhanced Convolutional Neural Network” Date of Publication: 26 April
2021 , Doi: 10.1109/ACCESS.2021.3075731
[3] Zhijian Yin, Boyang Wan , Feiniu yuan , Xue Xia and Jinting shi “A Deep Normalization and Convolutional Neural Network for Image Smoke Detection” Date of Publication: 30
August 2017 , DOI: 10.1109/ACCESS.2017.
[4] Xiwen chen, Bryce Hopkins , Hao Wang , Leo O’Neill, Fatemeh Afghah and Abolfazi “Wildland Fire Detection and Monitoring Using a Drone-Collected RGB/IR Image
Dataset” Date of Publication: 17 November 2022, DOI: 10.1109/ACCESS.2022.
[5] Jie Shi,Wei Wang , Yuanqi Gao, Nanpeng Yu “ Optimal Placement and Intelligent Smoke Detection Algorithm for Wildfire-Monitoring Cameras” Date of Publication: 15 April
2020, DOI: 10.1109/ACCESS.2020.2987991
[6] Gao Xu, Qixing Zhang, Dongcai Liu, Gaohua Lin and Jinjun Wang ”Adversarial Adaptation From Synthesis to Reality in Fast Detector for Smoke Detection” Date of
Publication: 04 March 2019, DOI: 10.1109/ ACCESS.2019.2902606
[7] Zhe Liu, Xiaoping Yang, Ying Liu and Zhihong Qian “Smoke-Detection Framework for High-Definition Video Using Fused Spatial- and FrequencyDomain Features ” Date of
Publication: 03 July 2019 , DOI: 10.1109/ACCESS.2019.2926571
[8] Feiniu Yuan, Xue Xia, Jinting Shi, Hongdi Li and Gang Li “Non-Linear Dimensionality Reduction and Gaussian Process Based Classification Method for Smoke Detection ”
Date of Publication: 25 April 2017 DOI:10.1109/ACCESS.2017.
[9] Yanbei Liu, Wen Qin, Kaihua Liu, Fang Zhang and Zhitao Xiao ”A Dual Convolution Network Using Dark Channel Prior for Image Smoke Classification” Date of Publication:
08 May 2019 ,DOI: 10.1109/ACCESS.2019.2915599
[10] Yinghao Gan, Weiqi Zhang, Jinfeng Zhangm “Building Fire Information Simulation Based on BIM Technology” Date of publication: 19 October 2023
DOI:10.1109/IECON51785.2023.10312378
ICNWC2024 19
