International Journal of Computing and Digital Systems

ISSN (2210-142X)
Int. J. Com. Dig. Sys. 14, No.1 (Aug-2023)
http://dx.doi.org/10.12785/ijcds/140141

Deep Learning-Based Real-Time Weapon Detection System

Amjed Al-Mousa1 , Omar Z. Alzaibaq1 and Yazan K. Abu Hashyeh2
1
Computer Engineering Department, PSUT, Amman, Jordan
2
Electrical Engineering Department, PSUT, Amman, Jordan

Received 4 Feb. 2023, Revised 18 Mar. 2023, Accepted 06 May. 2023, Published 01 Aug. 2023

Abstract: In recent years, the rate of gun violence has risen at a rapid pace. Most current security systems rely on human personnel
to monitor lobbies and halls constantly. With the advancement of machine learning and, specifically, deep learning techniques, future
closed-circuit TV (CCTV) and security systems should be able to detect threats and act upon this detection when needed.
This paper presents a security system architecture that uses deep learning and image-processing techniques for real-time weapon
detection. The system relies on processing a video feed to detect people carrying different types of weapons by periodically capturing
images from the video feed. These images are fed to a convolutional neural network (CNN). The CNN then decides if the image
contains a threat or not. If it is a threat, it would alert the security guards on a mobile application and send them an image of the
situation. The system was tested and achieved a testing accuracy of 92.5%. Also, it was able to complete the detection in as fast as 1.6
seconds.

Keywords: CNN, Security Cameras, Weapon Detection, CCTV, Deep Learning

1. Introduction [11]. They can also be used to make surveillance systems

The threat of mass shootings and robberies has been able to detect weaponized personnel and alert the proper
rampant in recent years. Moreover, the gun violence crime authorities immediately [12].
rates are very high, especially in countries where carrying
firearms is legal. In the US, the total number of gun The proposed architecture will impact the security
violence-related deaths in 2022 from January until the surveillance field in various aspects as it integrates tradi-
15th of September is 31,499 [1]. This is in addition to tional technology with the thriving areas of AI, ML, and
28,194 injuries, a massive number that cannot be ignored. the internet of things (IoT), [13], [14]. The most critical
Therefore, advanced modern security systems are essential expected impacts can be listed as follows:
to limit this problem.
One of the most widely used security systems is CCTV • Attacks are prevented before happening
cameras. CCTV is a type of situational crime prevention • Less successful robberies
(SCP) strategy in which levels of formal surveillance
are increased within a target area. SCP is focused on • Fewer casualties in incidents involving weapons and
preventing crime by reducing the number of criminal shooting
opportunities [2], [3]. These systems rely on a surveillance
operator monitoring multiple screens simultaneously, and • Security and surveillance running costs are reduced
while waiting for those unlikely-to-happen threats, they
become less focused, and detecting threats becomes harder This work’s core contribution is showing the ability
[4]. Therefore, these systems are not very efficient and are of DNNs in detecting weapons by developing a DNN
economically challenging because of the running costs of architecture and training it by creating a massive dataset
the surveillance operators. of images of people. Some people would be entering a
lobby without any weapons Negative, while others enter the
lobby carrying weapons; these will be tagged as Positive.
Artificial intelligence (AI) and machine learning (ML) These tagged images are then used to train a DNN that
techniques have been used successfully to solve many medi- is capable of classifying any image as a threat (Positive)
cal, technological, educational, and financial challenges [5], or not (Negative). Once the data is collected and the DNN
[6], [7], [8], [9]. Deep neural networks (DNNs) have also is trained, the system can be deployed in production. The
gained even more traction in image-based problems [10], proposed deployment can use existing indoor cameras to
E-mail address: [email protected], [email protected], [email protected] https:// journal.uob.edu.bh
532 Al-Mousa, et al.: Deep Learning-Based Real-Time Weapon Detection System.

monitor people entering a building lobby. The camera(s) The third approach by Verma and Dhillon uses Fast R-CNN,
should be installed facing the entrance area providing and deep learning [21]. A VGG-16-based classification
real-time images. Once the system detects the threat, it model (16 convolutional layers) is used, which focuses on
automatically notifies the security guards in the building on prediction loss minimization. Its methods are the same as
their mobile phones. The security guards receive an image the previous approaches, but with a minor difference of
of what is happening in that building area and respond having a fixed input size of 224p x 224p RGB images.
accordingly. It is worth noting that this work does not extend The fourth approach [22] uses the same basics mentioned
to concealed weapons [15], nor the detection of suspects above (sliding window, HOG feature extraction, SIFT,
based on worrying expressions or unnatural behavior [16]. and Harris key point detection) with unipolar sigmoid
It is also not concerned with predicting crimes or robberies and bipolar sigmoid as activation functions. The idea
before they happen [17]. is to detect the humans first and then check whether
there is a weapon in the picture. Notice that detecting
To summarize the objectives of the paper are as follows: humans using background reduction is faster than HOG,
which is essential in real-time applications. However, for
• Establish the ability of CNNs in detecting handheld background reduction to be efficient, the camera must
weapons using CCTV feeds. be installed indoors because the background subtraction
method is not flexible and can be affected by slight changes
• Identify the proper image pre-processing steps needed
in light intensity or object occlusions.
to facilitate the detection process.
The fifth approach by Lai and Maples tried various
• Achieve reasonable performance metrics using the pre-trained models such as VGG-16 mentioned earlier,
proposed architecture. Overfeat1, Overfeat2, and Overfeat3 [23]. The problem the
paper is trying to address is mainly detecting in real-time,
The rest of the paper is organized as follows: Section and despite the significant accuracy in GoogleNet Overfeat,
2 provides an overview of the literature and related work. it needed 14s per image, which is impractical. This paper’s
Section 3 details the method used to solve the threat detec- final results of reasonably high accuracy and a satisfactory
tion problem, including the details of the DNN model used. classification time of 1.3 seconds are achieved using
The results of the system testing are discussed in section Overfeat-3 with tuned hyper-parameters.
4. Section 5 presents the proposed system’s limitations The sixth approach proposes a detection model
and practical considerations. Finally, section 6 presents the (ResNet50V2) trained on the Open Images V6 dataset
work’s conclusion and future research directions. to detect the visual relationships of “holds” and “wears”
between people and objects [24].
2. Related Work and Literature Review The seventh approach utilizes the transfer learning concept
This section summarizes a few papers that used to train the VGGNet architecture based on VGGNet-16
different approaches in automatic weapon detection using weights [25]. While it achieved high accuracy, the dataset
deep learning and image processing. used is for stand-alone weapons. Another approach based
The first approach uses background reduction to remove on the VGGNet architecture classified seven classes of
static objects by using a reference image of the place, weapons and achieved an accuracy of 98.4 which was
then compares it to a present image of the same place to higher than other basic models like VGG16 and Resnet.
eliminate similarities [18]. By using Canny edge detection, The training was done on a small dataset from the internet
the obtained image will be a silhouette. After that, the containing just over five thousand images. However, the
model uses a sliding window, MPEG-7 feature extraction, images trained only contained the weapon with no other
and support vector machine (SVM) to decide whether to object. In addition, there was no mention of the time
send an alert. The sliding window size is determined by required to process each image [26].
trial and error after installing the system, which means the In[27], the main focus point was cold steel weapons.
system needs to be adjusted with different sliding window It tackled the challenge formed by the light reflection
sizes depending on installation. on this type of weapon by utilizing different regional
The second approach uses deep learning for automatic proposal search algorithms. And using Dacolt darkening
handgun detection [19]. It compares the sliding window and contrast to handle different brightness and lighting
results and Regional Convolutional Neural Networks (R- conditions.
CNN). After using The Histogram of Oriented Gradients
(HOG) descriptor for feature extraction and then using Based on the summary of the previous work, it is clear
a sliding window, the processing speed for detecting that there is a gap in finding a method that is:
pedestrians is 14 s/image which is not very practical
for real-time applications. This is in addition to the • Able to detect weapons in near real-time.
large computational power needed for the sizeable neural
• Computationally light.
network. The real challenge is to create a method that
dynamically optimizes all of the CNN’s parameters • Tested on more than one type of weapon.
simultaneously [20].
https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 14, No.1 531-540 (Aug-2023) 533

TABLE I. Dataset classes distribution

Tag Count Percentage

Positive 10k 36%
Negative 18k 64%

on how to hold the gun replicas. Some students pointed

them out, some held them on their side, while others
concealed or semi-concealed the guns. The images are
captured with a resolution of 1920p x 1080, a resolution
that can be scaled down if it turns out that a lower resolution
might be more suitable. The total size of the dataset is
Figure 1. High-level block diagram of the proposed system around 14k. However, data augmentation techniques, such
as scaling, rotation, and reflection, are used to get more
images. If an image and a horizontally reflected replica of
• Tested in a real environment with noisy image back- that image enter a neural network, it will process them as
ground. two completely different examples. Thus, from the captured
images, twice the original dataset size is achieved, around
• Auto-configured and does not need lots of calibration. 28k as in Table I. Note that the target here is to detect
whether there is a gun or not, not detect the gun replica
3. Methodology type.
Figure 1 shows a high-level block diagram of how the
system is expected to operate. The Raspberry Pi Camera A core problem in such datasets is that the target, in
captures the real-time video feed and feeds it to the process- most cases, represents a small object in the image, making
ing unit. The processing unit performs the image processing it harder to classify an image as a negative because any
steps required and then feeds the processed images to the shape close to the shape or the color intensity of guns
pre-trained DNN. The DNN was trained using a large might be classified as a positive. Figure 2 shows positive
number of images that were collected and properly labeled class examples on the top and negative class examples on
previously. The trained DNN acts as a classifier for the the bottom.
incoming image, whether it contains a weapon. If the DNN
is classified as Positive, i.e., contains a weapon, the system B. Image Processing
alerts the security guards on their smartphones and sends
them an image of the current situation. Otherwise, the image For the DNN to focus its learning on the important
is dropped, and the system proceeds to process a new frame. aspects of the image and not on static objects in the
The details of the system components are elaborated on in lobby. An empty reference image is captured as background
the following sections. and continuously subtracted from new images being taken.
Thus, static objects like furniture will be removed from
A. Experimental Setup & Dataset Collection the image. Initially, images were loaded with the resolution
Data is the most critical part of building any deep- (1920p x 1080p), which was slow to process. Therefore, the
learning solution. Such image-based applications require images are resized to (960p x 540p). Figure 3 summarizes
a large amount of data (images). Data is collected in the pre-processing steps applied to each input image.
the Innovation Lab at Princess Sumaya University for
The first stage is image smoothing. Gaussian Blur
Technology (PSUT). Raspberry Pi B+ (An ARM Cortex-
removes the salt and pepper noise. Figure 4 shows a
A53 1.4 GHz processor with WiFi capability and 1GB of
comparison between the image without smoothing and
SRAM) is used with a connected camera set on a table 1.8
with smoothing using both median and Gaussian blur.
meters high. The camera faces the lab entrance, which is
Gaussian blur is chosen. If one chooses not to blur
roughly 4.5 meters away. The Raspberry Pi is connected
the images as in Figure 4a, some unwanted edges of
to a display to monitor the process. A live log of the
the door and other components remain. Using Gaussian
time required to process and classify each image and the
blur is shown in Figure 4c, where all unwanted edges
classification results are displayed.
are filtered out. Meanwhile, the median blur is shown
in Figure 4d. Here some noise is filtered out, but the
The chosen area to collect the data is very active, edges of the door remain because median blur keeps the
with many students entering and exiting during the day, edges. After the mask is generated, the reference image is
which helps collect data for training and testing the model. subtracted, and a binary threshold is applied. Furthermore,
Students’ pictures were taken with their permission. In ad- morphological transformations are needed to filter noise
dition, students were given three types of handgun replicas after the threshold. The opening is used, which performs
to mimic an attack. Students were given little restriction erosion, then dilation. Iterating this procedure several

https:// journal.uob.edu.bh
534 Al-Mousa, et al.: Deep Learning-Based Real-Time Weapon Detection System.

Figure 2. Samples from the dataset

Figure 3. Flowchart presents a summary of image pre-processing steps

times removed most of the image’s noise. The kernel size mask to the original image. Figure 6 shows the final output
selected is (5, 5), and the number of iterations is two. after performing image processing steps. Figure 6a has too
Figure 5 shows the results of using too many iterations many features, which might confuse the model, increase
on a small image. Figure 5a shows the output after two the time needed for model training, and negatively affect
iterations, while Figure 5b shows the output after six the accuracy. In contrast, Figure 6b with the background
iterations. Figure 5b shows less noise; however, the critical eliminated and the human with or without the weapon is
information, which is the firearm held by the person, is visible, the desired features appear to be more evident than
cropped because of the excessive number of iterations. those in Figure 6a, and the system’s accuracy would be
Processing images in grayscale instead of colored images higher.
increases the speed, so the image is converted to grayscale,
and another binary threshold is applied. The most effective
way to apply the mask is using bitwise logical operations. C. The CNN Architecture
Bitwise XOR is used, which would result in inverting The input layer is the input interface of the neural net-
the foreground. Then a bitwise AND is used to apply the work where input images are loaded. Because of the GPU
memory limit and relatively small dataset size, a resolution
https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 14, No.1 531-540 (Aug-2023) 535

Figure 4. Applying blur. (a) Original image. (b) Subtracted without blurring. (c) Gaussian blur. (d) Median blur

Figure 5. Applying opening. (a) using two iterations. (b) using six iterations

of 640p x 360p (aspect ratio of 16:9) is chosen. Grayscale

images are used for training rather than colored images, as
the image colors do not add much useful information and
are simpler to process. Accordingly, the dimensions of the
input layer are (360, 640, 1).
After referring to [28] and experimenting with setting the
number of the hidden layers, three convolution hidden
layers are chosen with the ReLU activation function. The
number of input variables is the number of pixels, which
is 230,400. Based on [28], and [29], the number of feature
maps in the hidden layers is 64, 128, and 256 feature maps,
respectively. Flattening is applied to the third hidden layer
(a) The input image
so a fully connected layer can be added afterward with a
width of 2048 and a ReLU activation function.
After each of the three hidden layers, a pooling layer is
applied to the previous layer’s output. The pooling function
used is max-pooling as weapons appear as the brightest
objects in the pre-processed images as they are black in the
original images, and bright pixels carry higher pixel values
than dark pixels in the grayscale format, so when max-
pooling is applied, bright objects will be highlighted, and
dark objects will vanish as in Figure 7. Average pooling
does not highlight the desired features, but it is still better
(b) Background subtracted image than min-pooling, where white objects fade away. The size
of the max-pooling window applied after each hidden layer
Figure 6. Background subtraction
is 3x3.

https:// journal.uob.edu.bh
536 Al-Mousa, et al.: Deep Learning-Based Real-Time Weapon Detection System.

the DNN. The summary of its specifications is as follows:

• 2 vCPUs with a memory of 13GB.

• NVIDIA Tesla K80 with a memory of 12GB.

D. Alert Notification System

The Raspberry Pi and the mobile application are con-
nected using Wi-Fi technology in a local network. If the
system detects a threat, a push notification is sent to the
mobile if the application is not open to alert the user.
The application and the server are connected to Firebase.
Firebase acts like the middleman between the client and the
server, making it easier for the server to handle multiple
clients simultaneously. The notification and image are sent
to all the connected devices simultaneously. When the client
connects to the server, the server loads the resized captured
image, converts it into a byte array, and sends it. A smaller
version of the image is used here to send the image faster.
4. Results and Discussion
The results shown below will cover three architectures.
The first is the baseline architecture. While the second in-
vestigates the use of drop-outs. Finally, the third architecture
Figure 7. A comparison between different pooling functions [30] uses an additional layer to max-pool the inputs.
TABLE II. Detailed CNN architecture A. Baseline Architecture
Layer # Layer (Type) Output Shape Params # The first experiment uses the dataset without pre-
processing, where the collected data is used directly to
1 InputLayer 640 x 360 x 1 0
train the model. Here the training and testing accuracy is
2 Conv2D 638 x 358 x 64 640
between 57.8% - 62.5%, which is not adequate for the
3 MaxPooling2D 212 x 119 x 64 0
application to be successful. Thus, this approach is dropped,
4 Conv2D 210 x 117 x 128 73856
and pre-processing steps are applied to eliminate unneeded
5 MaxPooling2D 70 x 39 x 128 0
information and increase the training and testing accuracy.
6 Conv2D 68 x 37 x 256 295168
The dataset is shuffled and split into 90% for training
7 MaxPooling2D 22 x 12 x 256 0
and 10% for validation. Before splitting, 400 examples are
8 Flatten 67584 0
randomly removed from the dataset and stored in a separate
9 Dense 2048 138414080
folder for testing purposes. Although testing on such a
10 Dense 1 2049
small test dataset might not provide precise results, it is
chosen to be small to leave as much data for training as
possible. Batches of 32 labeled images are used to train
Finally, the output layer contains only one neuron
the CNN. After training the CNN designed earlier for ten
because the problem here is a binary classification problem
epochs, results are plotted as shown in Figure 8 showing
with two possible outputs that can be carried on a single
the evolution of the training accuracy, and the validation
neuron. The model structure is tabulated in Table II. The
accuracy as the number of epochs (iterations) increases. The
optimization algorithm for minimizing the cost function
training accuracy seems to increase much faster than the
is mini-batch gradient descent; it avoids the overshooting
testing accuracy; however, the model appears to continue
resulting from the high variance in the stochastic gradient
learning, and the accuracy is improving.
descent and guarantees convergence to the global minimum,
achieved in batch gradient descent[31]. The cost function After the tenth epoch, the model is tested on the testing
used is binary cross-entropy, and the learning rate has dataset (400 samples, 200 per class). The testing results are
been manually tweaked, starting from a typical value of shown in Table III. Accuracy, recall, precision, and f1-score
around 0.01 for multi-layer networks [32], concluding that are calculated using Equations 1-4:
the optimal learning performance has been observed at a
learning rate of 0.1.
T rue Positive + T rue Negative
Accuracy = (1)
A high-end virtual machine (VM) instance is customized T est Dataset S ize
on Google cloud platform (GCP) to perform the training of
https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 14, No.1 531-540 (Aug-2023) 537

B. Dropout Architecture
To reduce the gap between the training and validation
accuracy, ‘dropout’ is added after each of the hidden layers.
What dropout does is that it ‘drops out’ random neurons
at a specific rate as a regularization technique to reduce
overfitting. As a result, more training epochs are required
to reach the same training accuracy because some neurons
are dropped during training, but each epoch will take a
shorter time. The difference between the validation and
the training accuracy has decreased during the first few
epochs, as shown in Figure 9. However, results have not
improved, as the validation accuracy has not exceeded
91.9%. One might notice that stopping learning at the fifth
epoch might be better than continuing until the twelfth
epoch, as more overfitting occurred while not significantly
increasing the validation accuracy. This technique is one of
the regularization techniques used in machine learning, and
Figure 8. Training accuracy and validation accuracy for baseline it is called ‘early stopping’.
architecture

TABLE III. Performance metrics for baseline architecture

Metric Value
Validation Accuracy 92.4%
True Negative 187 out of 200
True Positive 183 out of 200
Accuracy 92.5%
Precision 93.4%
Recall 91.5%
F1-score 92.4%
Pre-Processing Time 0.3 s
Classification Time 2.0 s

T rue Positives
Precision = (2)
T rue Positives + False Positives Figure 9. Training accuracy and validation accuracy for dropout
architecture

T rue Positives
Recall = (3) C. Max-pooling Inputs Architecture
T rue Positives + False Negatives
A max-pooling function of 2x2 window size is applied
after the input layer and before the hidden layers to high-
2 ∗ Precision ∗ Recall light the desired features as a pre-processing step before
F1 − S core = (4)
Precision + Recall training. This enabled the increase of the batch size to 64
and has shown significantly faster learning and validation
Lower-resolution images (640p x 360p) drastically performance. The increase in the batch size can be justified
decreased the time required to pre-process and classify by the resolution reduction (the image entering the CNN
each image compared to the full-size images. The time now has a quarter of the number of pixels in the original
decreased from 9.5 seconds to 2.3 seconds, divided as image), which significantly reduced the trainable parameters
follows: the average pre-processing time is 0.3 seconds, from 139 million to 76 million. It has also given similar
and the average classification time is 2 seconds. This also validation and testing results: validation accuracy of 91.8%
decreased the time required to send the image, as the time and test accuracy of 92.5%, as shown in Table IV. Although
required to encode, transmit, and decode the image was this model has a slightly lower validation accuracy, it is
also drastically reduced. highly preferred over the last two architectures because of
its faster classification response. The average classification
time is now 1.3 seconds, down from 2.0 seconds in the base-
line architecture. These times were measured by embedding
https:// journal.uob.edu.bh
538 Al-Mousa, et al.: Deep Learning-Based Real-Time Weapon Detection System.

TABLE IV. Performance metrics for max-pooling inputs architecture

The current system was trained and tested on a reason-
Metric Value ably practical setup. However, future work can be expanded
Validation Accuracy 91.8% to analyze the limitations presented in section 5. In addition,
True Negative 185 out of 200 other new deep architectures, like fully convolutional neu-
True Positive 185 out of 200 ral networks, should also be evaluated. Such architectures
Accuracy 92.5% might enhance the detection accuracy while reducing the
Precision 92.5% required computations.
Recall 92.5%
References
F1-score 92.5%
Pre-Processing Time 0.3 s [1] GunViolenceArchive, “Gun violence archive.org,” 2022. [Online].
Available: https://www.gunviolencearchive.org
Classification Time 1.3 s
[2] E. L. Piza, B. C. Welsh, D. P. Farrington, and A. L. Thomas, “Cctv
surveillance for crime prevention, a 40-year systematic review with
timestamps within the code, thus accurately measuring the meta-analysis,” Criminology & Public Policy, vol. 18, no. 1, pp.
time from capturing the image till the classification is 135–159, 2019.
decided.
[3] K. L. Ritchie, D. White, R. S. S. Kramer, E. Noyes, R. Jenkins,
5. Limitations & Practical Considerations and A. M. Burton, “Enhancing cctv: Averages improve face identi-
While a fully functional demonstration has been pre- fication from poor-quality images,” Applied Cognitive Psychology,
vol. 32, no. 6, pp. 671–680, 2018.
sented in section 4, there are several factors to be considered
if such a system is to be deployed in a real environment: [4] R. Nunes-Vaz and S. Lord, “Designing physical security for com-
plex infrastructures,” International Journal of Critical Infrastructure
• Lighting: The system should be trained and tested Protection, vol. 7, no. 3, pp. 178–192, 2014.
to assess the impact of lighting on how resilient the
CNN architecture is to lighting changes. [5] H. Al-Zubaidi, M. Dweik, and A. Al-Mousa, “Stroke prediction us-
ing machine learning classification methods,” in 2022 International
• Weapon Type: The current dataset was created using Arab Conference on Information Technology (ACIT), 2022, pp. 1–8.
three types of handguns. The dataset can be expanded
[6] M. Atari and A. Al-Mousa, “A machine-learning based approach
to include other types of weapons, like knives, rifles, for detecting phishing urls,” in 2022 International Conference on
and semi-automatic guns. Intelligent Data Science Technologies and Applications (IDSTA),
2022, pp. 82–88.
• Camera Positioning: The results presented are based
on a fixed position selected to allow a frontal view [7] Z. Bitar and A. Al-Mousa, “Prediction of graduate admission using
of subjects. Further experiments with other positions multiple supervised machine learning models,” in 2020 Southeast-
should be considered to analyze the impact of the Con, 2020, pp. 1–6.
position on the system’s accuracy.
[8] S. Khalifeh and A. A. Al-Mousa, “A book recommender system
using collaborative filtering method,” in International Conference
• Image Quality & Specifications: While the system on Data Science, E-Learning and Information Systems 2021, ser.
had to down-sample the images from the camera DATA’21. New York, NY, USA: Association for Computing
feed, images with different resolutions, aspect ratios, Machinery, 2021, p. 131–135.
and those taken with non-standard lenses should be
analyzed. [9] A. Atwah and A. Al-Mousa, “Car accident severity classifica-
tion using machine learning,” in 2021 International Conference
6. Conclusion & Future Work on Innovation and Intelligence for Informatics, Computing, and
Technologies (3ICT), 2021, pp. 186–192.
The main goal of this research was to develop an effec-
tive architecture that can detect different types of weapons in [10] A. J. Moshayedi, A. S. Roy, A. Kolahdooz, and Y. Shuxin, “Deep
building entrances and alert security guards of any incidents. learning application pros and cons over algorithm,” EAI Endorsed
This was achieved by tapping into an image feed and Transactions on AI and Robotics, vol. 1, no. 1, p. e7, Feb.
processing captured images through a pipeline of image 2022. [Online]. Available: https://publications.eai.eu/index.php/airo/
article/view/19
pre-processing steps that reduce noise, reduce size, and
highlight the valuable information in the image. Processed [11] A. J. Moshayedi, A. S. Khan, S. Yang, and S. M. Zanjani, “Personal
images are later fed to a CNN. The CNN was trained image classifier based handy pipe defect recognizer (hpd): Design
using more than 28K labeled images. The baseline system and test,” in 2022 7th International Conference on Intelligent
achieved a 92.5% accuracy, 93.4% precision, 91.5% recall, Computing and Signal Processing (ICSP), 2022, pp. 1721–1728.
and 92.4% f1-score. All of these numbers have changed to
[12] A. J. Moshayedi, A. S. Roy, A. Taravet, L. Liao, J. Wu, and
92.5% with the max-pooling architecture. Also, it processed M. Gheisari, “A secure traffic police remote sensing approach via
and classified an image within 1.6 s. Thus, eliminating the a deep learning-based low-altitude vehicle speed detector through
need for continuous human monitoring and eliminating the uavs in smart cites: Algorithm, implementation and evaluation,”
chances of human errors.
https:// journal.uob.edu.bh
 Int. J. Com. Dig. Sys. 14, No.1 531-540 (Aug-2023) 539

Future Transportation, vol. 3, no. 1, pp. 189–209, 2023. [Online]. vol. 11, no. 16, 2021. [Online]. Available: https://www.mdpi.com/
Available: https://www.mdpi.com/2673-7590/3/1/12 2076-3417/11/16/7535

[13] Z. Ullah, F. Al-Turjman, L. Mostarda, and R. Gagliardi, [27] A. Castillo, S. Tabik, F. Pérez, R. Olmos, and F. Herrera, “Brightness
“Applications of artificial intelligence and machine learning in guided preprocessing for automatic cold steel weapon detection in
smart cities,” Computer Communications, vol. 154, pp. 313–323, surveillance videos with deep learning,” Neurocomputing, vol. 330,
2020. [Online]. Available: https://www.sciencedirect.com/science/ pp. 151–161, 2019. [Online]. Available: https://www.sciencedirect.
article/pii/S0140366419320821 com/science/article/pii/S0925231218313365

[14] Z. Sabeur, C. M. Angelopoulos, L. Collick, N. Chechina, [28] S. Karsoliya, “Approximating number of hidden layer neurons in
D. Cetinkaya, and A. Bruno, “Advanced cyber and physical situation multiple hidden layer bpnn architecture,” International Journal of
awareness in urban smart spaces,” in Advances in Neuroergonomics Engineering Trends and Technology, vol. 3, 2012.
and Cognitive Engineering, H. Ayaz, U. Asgher, and L. Paletta, Eds.
Cham: Springer International Publishing, 2021, pp. 428–441. [29] M. Madhiarasan and S. N. Deepa, “Comparative analysis on hidden
neurons estimation in multi layer perceptron neural networks for
[15] M. Parande and S. Soma, “Concealed weapon detection in a human wind speed forecasting,” Artificial Intelligence Review, vol. 48,
body by infrared imaging,” International Journal of Science and no. 4, pp. 449–471, Dec 2017.
Research (IJSR), vol. 4, pp. 182–188, 2015.
[30] M. Basavarajaiah, “Maxpooling vs minpooling vs average
[16] H. Bouma, J. van Rest, K. van Buul-Besseling, J. de Jong, and pooling,” 2021. [Online]. Available: https://medium.com/@bdhuma/
A. Havekes, “Integrated roadmap for the rapid finding and track- 95fb03f45a9
ing of people at large airports,” International Journal of Critical
Infrastructure Protection, vol. 12, pp. 61–74, 2016. [31] S. Ruder, “An overview of gradient descent optimization algo-
rithms,” 2017.
[17] M. P. de la Cruz López, J. J. Cartelle Barros, A. del Caño Gochi,
M. C. Garaboa Fernández, and J. Blanco Leis, “Assessing the risk [32] Y. Bengio, Practical Recommendations for Gradient-Based Train-
of robbery in bank branches to reduce impact on personnel,” Risk ing of Deep Architectures. Berlin, Heidelberg: Springer Berlin
Analysis, vol. n/a, no. n/a, 2021. Heidelberg, 2012, pp. 437–478.

[18] M. Grega, A. Matiolański, P. Guzik, and M. Leszczuk, “Automated

detection of firearms and knives in a cctv image,” Sensors, vol. 16,
no. 1, 2016.

[19] R. Olmos, S. Tabik, and F. Herrera, “Automatic handgun detection

alarm in videos using deep learning,” Neurocomputing, vol. 275, pp.
66–72, 2018.

[20] M. S. Al-Daweri, S. Abdullah, and K. A. Z. Ariffin, “A homoge-

neous ensemble based dynamic artificial neural network for solving
the intrusion detection problem,” International Journal of Critical Amjed Al-Mousa a senior IEEE member
Infrastructure Protection, vol. 34, p. 100449, 2021. and an associate professor of computer en-
gineering at PSUT since 2012. Dr. Al-Mousa
[21] G. K. Verma and A. Dhillon, “A handheld gun detection using faster has received his Ph.D. from Santa Clara Uni-
r-cnn deep learning,” in Proceedings of the 7th International Con- versity, M.Sc. from Virginia Tech, and B.Sc.
ference on Computer and Communication Technology, ser. ICCCT-
2017. New York, NY, USA: Association for Computing Machinery, from the University of Jordan all in Elec-
2017, p. 84–88. trical Engineering. Dr. Al-Mousa combines
this academic experience with more than
[22] R. Vajhala, R. Maddineni, and P. R. Yeruva, “Weapon detection 13 years of industry experience at Silicon
in surveillance camera images,” Master’s thesis, , Department of Valley. He started his career by working at
Applied Signal Processing, 2016. Intel Corporation in 2001 and later moved to PDF solutions to
work on Design for Manufacturing software tools. After that, he
[23] J. Lai and S. Maples, “Developing a real-time gun detection classi-
fier,” http://cs231n.stanford.edu/reports/2017/pdfs/716.pdf, 2017.
assumed a senior manager position for data analytics at SEVEN
networks. Dr. Al-Mousa has served twice as the head of the
[24] T. Truong and S. Yanushkevich, “Detecting subject-weapon visual computer engineering department at PSUT. He developed course
relationships,” in 2020 IEEE Symposium Series on Computational content in the fields of Artificial Intelligence, Machine Learning,
Intelligence (SSCI), 2020, pp. 2047–2052. Cloud Computing, and Big Data. His research interests are in the
fields of intelligent systems design and machine learning with
[25] N. Dwivedi, D. K. Singh, and D. S. Kushwaha, “Weapon classi- varying industrial applications. He can be contacted at email:
fication using deep convolutional neural network,” in 2019 IEEE
[email protected].
Conference on Information and Communication Technology, 2019,
pp. 1–5.

[26] V. Kaya, S. Tuncer, and A. Baran, “Detection and classification

of different weapon types using deep learning,” Applied Sciences,

https:// journal.uob.edu.bh
540 Al-Mousa, et al.: Deep Learning-Based Real-Time Weapon Detection System.

Omar Z. Alzaibaq received his B.Sc. de- Yazan Abu Hashyeh received his B.Sc.
gree in Computer Engineering from Princess degree in electronics Engineering from
Sumaya University for Technology (PSUT), Princess Sumaya University for Technology
Jordan, in 2019. He is currently working as (PSUT), Jordan, in 2019. Since October
a systems and software engineer at Iotistic 2019, he has worked as an Artificial in-
Solutions. His main fields include develop- telligence engineer at IOTISTIC Solutions.
ment, Linux-based scripting, fleet manage- During his position as an Artificial Intelli-
ment systems, and artificial intelligence. gence engineer. He works part-time Lecturer
with Pioneers Academy teaching Python and
Artificial intelligence. His main field of work
is in image processing and software development.

https:// journal.uob.edu.bh