DEEP LEARNING BASED FIRE DETECTION,

2023 2nd International Conference on Vision Towards Emerging Trends in Communication and Networking Technologies (ViTECoN) | 979‐8‐3503‐4798‐2/23/$31.00 ©2023 IEEE | DOI: 10.1109/VITECON58111.2023.10157321

ALERT AND SUPPRESSION SYSTEM

N. Senthilnathan, Sibikkrishna N,
Professor, Department of Electrical and Electronics UG Scholar, Department of Electrical and Electronics
Engineering, Kongu Engineering College, Perundurai, Erode, Engineering, Kongu Engineering College, Perundurai, Erode,
India. India.
[email protected] [email protected]

Sowbarn T S, Niranjan R S,
UG Scholar, Department of Electrical and Electronics UG Scholar, Department of Electrical and Electronics
Engineering Kongu Engineering College, Perundurai, Erode, Engineering, Kongu Engineering College, Perundurai, Erode,
India. India.
[email protected] [email protected]

Abstract—The biggest and most dangerous threat to any filling the gap of previous systems. For the economy,
enterprise is fire. The primary causes of fire include static product management, and especially the cotton sectors, it
electricity, overheating, human mistake, environmental poses serious problems and a threat. Every building and
factors, and equipment failure. In this project, issues with public transportation passenger vehicle now has fire
current methods for detecting fire in industrial settings have
protection and fire prevention systems due to the steadily
been highlighted, and an original approach to address the
issues has been put forth, thereby offering functions like rising number of fire incidents. These systems frequently
alarm, detection, and suppression. This project uses an use point-type thermal and smoke detectors, which must be
Arduino-powered fire detector and a security camera, and it positioned near the fire for best performance in detecting
applies the ideas of image processing and deep learning to fires. These devices also need to be fitted and placed
produce results for real-time detection. Convolutional Neural correctly because of the fire risk. The process of
Network (CNN)/MobileNet-v2 is used in this research to functioning like computer vision, and AI is widely based
detect the proper image of fire and perform image pre- on usage of AI technique [1]. The testing data and Training
processing. There have been numerous large-scale fires and data are trained and to build CNN model. By using Adam
bad fire imagery in a lot’s of circumstance, having both
optimizer instead of SGD optimizer accuracy can be
interior (e.g., homes, offices, factories) and external (e.g.,
depository or cache areas). A low-cost embedded device improved from 62.5% to 95%. This accuracy results in
(Jetson Nano or raspberry pie) that consists of a single, fixed predicting fire every time without error. These devices may
camera for individual backdrop and functions in the seeable perform better than other well-known techniques in terms
spectral range can also be used to implement this job. The of response time and detection range.
video camera used for surveillance doesn't need to meet any
particular specifications. The camera will continue to keep an Conventional photometric, thermal, or chemical
eye on the surroundings of the room so long as the device is smoke/fire detectors can react in a matter of minutes but
activated. If a fire is found in the frame, CNN will process the need a huge fire or smoke to exist before an alarm can
fire image before processing the alert message and sending it sound. They also don't work well in open spaces. By filling
via email to the user's smartphone, alarm sound, and water the gap left by prior systems, new camera-based solutions
pump to put out the fire. are being created, which increase the resilience and
Keywords—Deep Learning, Convolutional Neural Network dependability of fire and smoke detection. The bulk of
(CNN), Image Processing, Surveillance, Water Pump. human settings, including city streets, commercial
buildings, and public transportation, are currently
I. INTRODUCTION monitored by cameras [2] and closed-circuit television
Every industry is most at risk from fire, which (CCTV) systems. A control room already has many
cannot be avoided. According to ADSI data, 11,037 fire cameras, a network for communication, potential
accidents were registered in India in 2021. Industry was processing units, and screens. Because they don't require
present in 34.3% of fire occurrences, the data shows. Fire any new items, using them would result in cost savings for
detection poses a serious threat to the majority of industrial both installation and purchasing costs. By including
sectors and presents significant logistical and financial additional software, the fire detection method can be
challenges. Most industries, including those that produce readily implemented into this system. It might be feasible
paper, cotton, textiles, crackers, etc., are susceptible to fires to implement a low-cost Internet of Things (IoT) system
[13]. Traditional smoke/fire sensors based on photometry, using distributed video camera nodes in an ad-hoc
thermal, or chemical detection can react within several architecture [15]. In addition to offering a web-based
minutes, requiring a large amount of fire/smoke to trigger streaming video platform, this Internet of Things (IoT)
an alarm. Moreover, they cannot work for outdoor scenes. distribution system might be able to instantly trigger a fire
It improves speed and accuracy of the fire detection. The alarm.
development of new camera-based solutions improves the
robustness and reliability of smoke and fire detection by

979-8-3503-4798-2/23/$31.00 ©2023 IEEE

 These detection systems were proposed and adopted
in the cotton sector thanks to the quick advancement of
digital camera and computer vision technologies. Initially,
manually produced features for identifying fire in a video
feed were extracted using the color and shape properties of
smoke. Artificial intelligence techniques have been
developed as a result of the increased use of high-
performance CPUs and graphics processing units (GPUs)
in recent years. In specialized computer vision applications
like photo categorization, deep learning models surpass
humans, significantly improving visual object recognition,
object detection, and many other domains. the ability of a
deep learning model to integrate classification and feature
extraction into a single network. Due to their ability to Fig 1. Early flame detection in surveillance video using deep
automatically extract data from unprocessed images, deep CNN.
learning algorithms have gradually supplanted the III. CONVOLUTIONAL NEURAL
historically developed visual identification methods [3].
NETWORK ARCHITECTURE
This study describes A fire detector device using
Arduino and surveillance camera has been created and the The CNN deep learning architecture is based on how
concept of Image Processing and Deep learning has been living things see their environment visually [4].
applied in order to get the real-time detection results. A
A CNN CNNs were created primarily to handle
large scale of fire/smoke and negative videos in different
pixel input and are essentially a class of artificial neural
environments, both indoor or outdoor has been used. The
networks used for recognition and image processing. Being
dataset was divided into two classes or labels. These were
able to internalize two-dimensional images through CNNs
‘fire’ and ‘without fire’, and together, the images were
is advantageous. Working with photos enables the model to
curated, aggregating to around 4000 images. Convolutional
learn scale and location in a range of data types. Every 2D
Neural Network (CNN)/MobileNet-v2 which is used for
or 3D array of data can then be subjected to it as necessary.
the pre-processing of image, to detect the correct image of
A particular sort of deep neural network called a CNN is
fire. Local surrounding will be detected continuously using
frequently used in picture categorization and computer
surveillance system. The video from real time surveillance
vision procedures. Figure 2 illustrates the main
system is separated into frame by frame. If surveillance
characteristics of a typical CNN design.
camera captures image video Arduino process that data
results in the fire detection, alarm and suppression. This
project made use of an Arduino-made detector.
II. THE PROPOSED FRAMEWORK
The Most of the research on conventional flame
detection methods for extracting features has been
conducted in the previous ten years. These techniques' main
drawbacks are their cumbersome feature engineering
process and poor flame detection performance.
Nonetheless, these techniques may result in erroneous
Fig 2. Main operation of a typical CNN architecture.
warnings and alerts, particularly when keeping an eye on
dark areas, altering lighting, and red objects. We examined
deep learning methods for early flame detection in-depth to The primary components of the design include
find solutions to these problems. We investigated a number fully connected layers, pooling layers, activation functions,
of CNNs in light of current advancements in integrated and 2D convolutional layers (conv2D). The suggested
developing abilities and the implicit for horizontal extent in model has five Conv2D layers with padding set to "same"
order to increase the accuracy of flame descrying and and a stride of one. Each conv2D layer recovers the feature
decrease the frequency of untrue alarms. Figure 1 illustrates map of the 2D input data by "sliding input" through a filter
the theoretical basis of our suggested method for flame or kernel, and then does the following: In the
detection in CCTV surveillance networks. aforementioned example, P is the input image's matrix and
Q is the convolutional kernel that results in C as the output.

C(Z)= (PQ) (x)=(P(Z)Q(Zx) dZ) (1)

Pooling layers reduces the amount of the feature map, as representation of a Haar kernel may be seen in the
seen in figure 3. Since there are fewer trainable parameters, rectangular area in the center, which only has pixels that are
calculations can be performed more quickly without either light or dark. The average pixel values at the brighter
sacrificing essential components. Maximal and average region and the darker zone are subtracted to perform the
pooling are the two basic pooling methods that can be Haar computation. The Haar characteristic is advantageous
applied [5]. when the difference is almost equal to 1.
This serves as an illustration of how to determine
a Haar value from a rectangular image slice. The Haar
feature's darker portions are reflected by pixels with values
of 1 and its lighter regions by pixels with values of 0.
Finding a particular component of the image is the
responsibility of each of these. any area of the image where
there is a noticeable shift in intensity, such as a line,
structure, or an edge. In the image above, for example, the
Haar feature might be able to identify a vertical border that
Fig 3. Convolution Pooling SoftMax has lighter pixels on its left and darker pixels on its right.
The objective in this situation is to determine what
The practise of "max pooling" involves putting the most
distinguishes "the sum of all image pixels lying in the
crucial data just above the kernel. The mean for each value
in the region is determined using average pooling, in lighter section of the Haar feature from the total of all image
contrast. pixels located in the darker area of the Haar feature." The
Haar value will be closer to 1 if there is an edge in the
picture separating the light pixels on the left from the dark
The conclusion or the centre of neural networks may
pixels on the right. This means that if the Haar value is near
have nodes with activation functions (layers). They have
1, we should declare the presence of an edge. As the Haar
the power to control a neuron's ability to fire. The hidden
and output layers need to have the right activation functions value [8] is far from 1, there is no edge in the definite.
in order to train a model effectively. ReLU, which is A vertical edge that divides is one example of the
frequently employed for hidden layers, is replaced by Haar property. For any extra image structures or edges with
SoftMax for the output layer. The probability distribution different directional runs, other Haar characteristics can
is produced via SoftMax using a real number vector. For already be applied. The Haar feature needs to look at the
addressing scenarios requiring a variety of classification full image to detect an edge [12].
classes, the latter option is favoured [6].

The Fully Connected (FC) layers are employed

following the use of all convolutional layers. Building
binary and multiclass categories for photographs is made
feasible by these layers. The SoftMax activation function is
the greatest option for obtaining probabilistic results in
these layers.

ReLU : f(x)=max (0, x) (2)

SoftMax :(xi) = exp(xi)/∑j exp(xj) (3)
A convolution method can generate feature maps
by merely applying a number of kernels of different sizes Fig 4. Haar cascade algorithm
to the input data. These feature maps serve as the input for
the subsequent subsampling or pooling procedure, which V. HARDWARE SETUP
chooses the greatest activations from them within a The output of the Convolutional Neural Network model is
predetermined region. These techniques are important for what the system utilizes to interpret its environment, as was
lowering the feature vectors' dimension and, to some already established. The relay is activated when a nearby
extent, establishing translation invariance. The CNN fire is discovered. The electrical components connected
pipeline also includes the fully connected layer, which across receive the power supply as soon as the relay is
models high-level abstractions from the input data [7]. The turned on. The fire is extinguished by the water suppression
convolution and fully connected layers in these three major system, which pumps water [14]. A simultaneous alarm and
operations are built up of neurons whose weights are letter to the appropriate person will be generated to track
trained on and modified to better represent the input data the progress of the procedure. Until the entire surrounding
over the course of training. region is consumed by flames, the process continues. The
IV. HAAR CASCADE ALGORITHM hardware configuration for the aforementioned technique is
shown in Figure 5 [10].
A sample image with pixel values ranging from 0.0 to
1.0 is shown in the left rectangle of Figure 4. A
 improperly managed is fire [11]. Hence, the creation of
early fire detection systems is crucial. We then offer a fire
detection neural network for surveillance footage in this
paper that is both practical and marketable. The model's
development was influenced by the Google Net
architecture, which put a premium on computational
complexity and detection precision [9]. Experiments have
revealed that the suggested design beats both the Alex Net
architecture-based fire detection approach as well as the
currently in use hand-crafted features-based fire detection
approaches. Despite the fact that our work improved the
precision of flame detection, many false alarms still occur,
prompting further study in this field. Additionally, it is
possible to discern between flames and smoke by
intelligently modifying the current flame detection

Fig 5. Hardware Setup of the project

VI. CIRCUIT DIAGRAM

An Arduino UNO that is fed the program
coding and attached to the necessary port for the operation
has the buzzer, relay, and water suppression pump properly
connected. The circuit diagram shown in Figure 6 depicts
how the circuit is set up overall. When a fire is detected, the
Arduino UNO activates the alert buzzer via the relay
system and tells the motor system to pump and suppress the
water.

frameworks. This will enable more challenging real-world

scenarios to be handled by video surveillance systems.
Fig 7. Hardware connection of the project

IX. REFERENCES
[1] Janku P., Kominkova Oplatkova Z., Dulik T., Snopek P. and Liba
J. 2018. “Fire Detection in Video Stream by Using Simple Artificial
Neural”. Network. MENDEL. 24, 2 (Dec. 2018), 55–60.
[2] Deep Learning-Based Fire Detection System with Thermal
Camera for Indoor Environments by Choi et al. (2019).
[3]. Seebamrungsat, Jareerat et al. “Fire detection in the buildings
Fig 6. Circuit Diagram of the project using image processing.” 2014 Third ICT International Student Project
Conference (ICTISPC) (2014): 95-98.
VII. HARDWARE CONNECTION [4] "Fire Detection Based on Convolutional Neural Networks Using
Thermal Images" by Yu-Xiang Zhang, et al. in IEEE Access. M. Ligang,
The relay, which is connected to the motor and alarm C. Yanjun and W. Aizhong, “Flame region detection using color and
buzzer, is attached to the port of the Arduino UNO, as motion features in video sequences,” The 26th Chinese Control and
shown in Figure 7. The transformer in the prototype Decision Conference (2014 CCDC), Changsha, 2014, pp. 3005- 3009.
configuration changes a 240 V source into a 9 V output. [5] Deep Learning-Based Fire Detection Using RGB Cameras by M.
The Arduino UNO will give the relay the command to A. Aboulhassan, A. Elshafei, and M. A. Eltawil. IEEE Sensors Journal,
activate when it detects a nearby fire. A motor connected vol. 21, no. 12, pp. 13170-13180, Jun. 2021.
through the relay engages when the relay operates, [6] A Deep Learning-Based Fire Detection System for Industrial
suppressing the water. The suppression's objective is to Settings by C. Yao, J. Wang, and X. Chen. IEEE Transactions on
Industrial Electronics, vol.68, no. 6, pp. 5131-5140, Jun. 2021.
extinguish fires in commercial environments. Both the
alarm buzzer and the postal announcement are on. [7] Real-Time Fire Detection System Using Deep Learning on
Thermal Images by M. Gharavi-Alkhansari, R. Farnoudpour, and M.
VIII. CONCLUSION Masoudi. IEEE Sensors Journal, vol. 21, no. 8, pp. 9976-9983, Apr. 2021.
[8] Fire Detection Using Deep Learning: A Review by S. Zia-Ur-
The ability of surveillance systems to recognize Rehman, A. Javaid, and M. Awais. IEEE Access, vol. 9, pp. 41592-41605,
Feb. 2021.
multiple anomalous events, such as fires, accidents, and
other emergencies, has lately grown, as has the processing [9] Fire Detection and Alert System Using Deep Learning by M. R.
power of smart devices. One of the potentially calamitous Kim, S. J. Kim, and Y. J. Kim. Electronics, vol. 9, no. 2, pp. 285, Feb.
2020.
occurrences that could cause significant losses if
 [10] A Deep Learning-Based Fire Suppression System for Smart
Buildings by S. C. Kim, C. Y. Kim, and Y. J. Kim. Sensors, vol. 20, no.17,
pp. 4813, Aug. 2020.
[11] Arif, M., K. Alghamdi, S. Sahel, S. Alosaimi, M. Alsahaft, M.
Alharthi, and M. Arif. 2021. Role of machine learning algorithms in forest
fire management: a literature review. J Robotics Autom, pp. 212–226.
[12] Wiame Benzekri1, A. E. (2020). (IJACSA) International Journal
of Advanced Computer Science and Applications, Vol. 11. Early Forest
Fire Detection System using Wireless Sensor Network and Deep
Learning.
[13] Ayala, A., Fernandes, B., Cruz, F., Macêdo, D., Oliveira, A. L.,&
Zanchettin, C. (2020, July).Kutralnet: A portable deep learning model for
fire recognition. In 2020 International Joint Conference on Neural
Networks (IJCNN) (pp. 1-8).IEEE.
[14] Jadon, A., Varshney, A., & Ansari, M. S. (2020).Low-
complexity high-performance deep learning model for real-time low-cost
embedded fire detection systems. Procedia Computer Science, 171, 418-
426.
[15] Fernandes, A. M., Utkin, A. B., & Chaves, P. (2022). Automatic
Early Detection of Wildfire Smoke With Visible Light Cameras Using
Deep Learning and Visual Explanation. IEEE Access, 10, 12814-12828.