Volume 10, Issue 9, September – 2025 International Journal of Innovative Science and Research Technology

ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25sep396

Garbage Detection Using Deep Learning Methods

(GD-DLM)
M. Mudasar Azeem1; Syed Anwaar Mehdi2; Muhammad Ali Shahid3;
Mubasher Hussain4; Muhammad Adnan5; Spogmai6; Bilal Shabbir Qaisar7*
1,4,5,6
Faculty of Computing University of Okara 56300, Pakistan
2,3,7
Faculty of Comsats University Islamabad, Sahiwal Campus 57000, Pakistan

Corresponding Author: Bilal Shabbir Qaisar7*

Publication Date: 2025/09/16

Abstract: In today’s expanding and densely populated world, it’s crucial to design an automatic intelligent garbage sorter
machine that uses advanced sensors. Garbage picture classification is a fundamental computer vision problem that must
be solved before sensors can be included in this system. This research presents a model for autonomous trash classification
using deep learning that can be applied in high-tech garbage sorting equipment. The 2,527 photos in the rubbish dataset
are categorized into six types: trash, cardboard, glass, metal, paper, and plastic. The next step is the creation of GD-DLM,
a deep learning model for garbage categorization that is an upgrade from Xception and DenseNet121 models. At last, the
tests are run to evaluate GD-DLM against the best-of-breed approaches to garbage classification. The suggested Xception
and DenseNet-121 models scored 92.11% and 88.63%, respectively, compared to the baseline accuracy.

Keywords: Machine Learning; Resnet50v2; Trauma; Medical Images.

How to Cite: M. Mudasar Azeem; Syed Anwaar Mehdi; Muhammad Ali Shahid; Mubasher Hussain; Muhammad Adnan;
Spogmai; Bilal Shabbir Qaisar (2025) Garbage Detection Using Deep Learning Methods (GD-DLM). International
Journal of Innovative Science and Research Technology, 10 (9), 608-618.
https://doi.org/10.38124/ijisrt/25sep396

I. INTRODUCTION cleanliness on the streets. In 2015, Imteaj developed an

application for the city of Dhaka, which is the capital of
The natural and anthropogenic sources of waste that Bangladesh, that was built on the Android operating system.
end up on urban roadway surfaces include things like leaves, It is possible for the user to make a contribution to the
dirt, silt, litter, and illicit dumping. Ineffective street cleaning cleanliness of his city by either contacting the city
has the potential to damage the city's economy, reputation, corporation or signalling to volunteers to come forward. As
and tourism. It has also been acknowledged that unclean of the year 2015, the city of Los Angeles had established a
roadways may significantly contribute to water and air cutting-edge method for evaluating the cleanliness of each
pollution [14, 20]. Studies have demonstrated that individual street. A map depicting the cleanliness of each and
individuals are more prone to litter when they see litter on the every one of Los Angeles' blocks was created for the first
streets. People are more likely to think twice before throwing time by the city. With the help of this new tool, Sanitation is
anything and ultimately refrain from littering if the streets are better able to target locations that have a high demand for
clean. Keeping city streets clean is essential since it affects services and to ensure that services are distributed
not just the city's reputation and image but also the quality of consistently. The majority of the technologies that are now in
life for city dwellers and employees. use for identifying trash are not entirely automated and still
rely on human intervention. This is despite the fact that these
Numerous municipalities have implemented a variety methods provide fresh ideas for communities to clean their
of strategies and exerted a substantial amount of work in streets. Therefore, in order to assess whether or not the
order to enhance the cleanliness of their streets. As an roadway is dirty, the cleanup crews will need to manually
illustration, New York public utilised a program known as collect and identify each photo. For this reason, a system that
Scorecard for the purpose of evaluating the cleanliness of its is both promising and ideal should be able to detect trash in
public streets and sidewalks. A stringent photographic each collected image in an autonomous and reliable manner
standard of cleanliness for streets and pavements serves as without the need for human interaction.
the basis for the measures. An index of cleanliness was
presented by Sevilla in 2013 for the city of Granada, which is Convolutional neural network (CNN) methods [1] are
located in southern Spain, in order to quantify the level of currently the norm for state-of-the-art object identification

IJISRT25SEP396 www.ijisrt.com 608

Volume 10, Issue 9, September – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25sep396

and recognition systems; these systems rely on deep learning technique would be suggested for the improvement
architectures like VGG. Such solutions work, but they are of accuracy. Following the main contributions of this
computationally expensive in processor, memory, and research:
storage space requirements. Hence, in many applications
with computational limits, like mobile devices, their  Developed a Garbage Detection System using Deep
utilisation is not practical. The usage of "mobile" CNN Learning Methods (GD-DLM) that recognizes the
architectures—lightweight CNN designs tailored for mobile cardboard, glass, metal, paper, plastic, and trash classes.
devices—is one viable alternative that has been investigated
in the literature. But as far as we are aware, no research has  Improved the accuracy of the existing deep learning
yet looked at how to evaluate appropriate litter detectors with model.
neural network topologies, especially in cases when devices
have limited processing power. Also, there are a lot of cloud II. LITERATURE REVIEW
computing services for machine learning, but they don't have
fast enough reaction times for real-time applications. This is A number of outstanding individuals have made
particularly true in outdoor areas like forests and beaches, indelible marks in the field of waste management-related
where network connection is spotty at best. In addition, there machine learning and internet of things (IoT). An image
is an increase in power consumption during data transfer via processing and Convolutional Neural Network (CNN)
the network. Thus, adopting lightweight neural networks has garbage categorisation system was built in a later study [11]
several benefits, such as reduced energy usage, local (Bobulski and Kubanek, 2019). Their research has been
processing, and efficiency. focused solely on detecting polyethylene. A battery of tests
was run by the researchers to determine the presence of
Mynation has lately implemented policies to reduce its terephthalate, polyethylene, polypropylene, and polystyrene.
own waste output in light of the concerning global increase Capsule-Net, a neural network that can differentiate between
in trash creation over the past few years. A domestic waste plastic and non-plastic products, was employed by the
management system for controlled release, recycling, scientists for solid waste management (Sreelakshmi et al.,
transportation, and treatment of waste is now mandated by 2019). Two publicly available datasets were determined to
the most recent round of revisions to the "Law on Preventing have an accuracy of 96.3% and 95.7%, respectively, by the
and Controlling Environmental Pollution by Fixed Wastes" writers. The full integration was built and tested using
of the People's Republic of China, which took place in 2020. multiple pieces of hardware.
Currently, the majority of the public environment's trash
sorting occurs in a limited number of designated areas. In a recent work (Huiyu and O, O. G. Kim, S. H.,
Ineffective sorting is just one of several issues; another is the 2019), the author developed a novel garbage categorisation
unpleasant working environment. It may be as simple as model by utilising deep learning processes. This method was
sorting trash at home to get to the core of the issue. Landfills also used for recycling waste. Adedeji and Wang (2019) put
collect an overwhelming variety of waste products, and it's out the idea that a deep learning model may be used to
hard to keep track of them all since individuals lack a basic automatically identify trash. The authors also mentioned that
understanding of classification and hence seldom separate the model was employed for the purpose of sorting
waste into discrete groups. A lot of people have been recyclable waste. The authors of the article state that in order
thinking about getting home care robots recently. One to classify trash, they utilised a Support Vector Machine and
example is the rapid industrialisation and widespread public a pre-trained Convolutional Neural Network (CNN) called
adoption of sweeping robots, which were among the first ResNet-50 (Nowakowski and Pamua, 2020). Comparing the
products of their kind. Concurrent with this, mass production model to a publically available dataset revealed an accuracy
of automobiles began. A lot of the sweeping robots on the rate of 87%. A recent study looked into a new method of
market now aren't very smart, even though they have some identifying and categorising e-waste, or electronic trash
basic features like path planning, autonomous charging, and (Misra et al., 2018). Classification was carried out using the
obstacle avoidance. CNN model, while the identification of different types of e-
waste was accomplished using the RCNN model. The
Our top goal should be maintaining a clean, healthy researchers discovered an accuracy range of 90% to 97% for
environment for future generations. Several issues with both detection and classification. Using the IoT for trash
employing deep learning for garbage detection have been management was not mentioned in any of the articles [11, 21,
raised in the current research. One problem was that several 23, 19, 24].
categories, including but not limited to cardboard, cardboard,
glass, metal, paper, plastic, and trash, had yet to be included An essential approach to automated, resilient waste
in earlier research. This research aims to address these gaps management was presented in a publication [26]. The authors
in the existing literature. demonstrated a smart garbage can using the ultrasonic sensor
in conjunction with a number of gas sensors. Additionally,
The current study proposes a Deep Learning Method they suggested using a cloud server and an Android app to
for Garbage Detection (GD DLM), which would identify the provide a live feed of trash. Here, ML methods were
dataset. The proposed method would be trained on the completely disregarded. A clever and economical waste
garbage dataset, which contains six classes such as management system was concocted by the author of a paper
cardboard, glass, metal, paper, plastic, and trash. A deep [4, 41]. As a result of the Internet of Things, the authors have

IJISRT25SEP396 www.ijisrt.com 609

Volume 10, Issue 9, September – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25sep396

integrated a number of devices, including a GSM module, an III. METHODOLOGY

ultrasonic sensor, and an Arduino small. When the trash level
rose above a certain point, the GSM module would notify the  Materials and Methods
user by text message. Each user received an audio message Self-teaching algorithms are crucial to the development
after the system was upgraded with a PIR motion sensor and of AI. Algorithms of this type can change and improve as
a memory card. Author also claimed sufficient performance more data is collected about a project [35]. The result of
from the proposed system. In this study, the authors [19] solutions-oriented technology is ongoing. For self-learning
created a method for intelligent waste management in urban software to function, it must have access to these
areas. We came up with a model that was efficient and cost- conceptualizations [7]. In the same way that genuine neural
effective. You may get some information about it in the networks are structured into layers, artificial neural networks
article [6]. (ANNs) similarly connect their nodes (neurons). This neural
network stores information, processes algorithms (with the
The authors have created a model that optimises weighting of either plus or minus), and provides sensory
garbage collection using the Internet of Things. An infrared feedback. The layered architecture and attention to detail of
sensor and a Raspberry pi form its basis. Better results from ANNs show promising results. These networks are capable
the garbage collection procedure may be achieved if the of "deep learning" [27, 28].
system management rescheduled and redirected the waste.
According to a report [9], someone built an intelligent trash In this research, we build a deep transfer learning
can that could measure the amount of trash. All of the system for waste classification. After the problem of class
proposed models made use of wireless networking and a web imbalance in the dataset has been fixed, various
server. The scientists determined the amount of trash still in augmentation methods are utilised to generate a large
the bin using an infrared distance measuring device. A quantity of additional data. In the second stage, attributes are
website transmitted these findings and data bits to an automatically extracted, and pre-trained algorithms for
Android app. Using IoT and a Raspberry Pi, the authors of a recognising and classifying rubbish are used. Figure 1 is a
paper [18] built an intelligent trash can. The paper's authors flowchart depiction of the suggested procedure.
found that by using an IoT-based smart garbage system, they
were able to decrease food waste. The designers integrated
all the parts into a unified whole using mesh technology. A
server and router were the components of the suggested
model for collecting and analysing data on food poisoning.
There was a 33% reduction in food waste as a result of the
numerous successful trials that were carried out. While there
were a number of articles that discussed trash management
[31, 22, 37], no one had actually laid out a plan for the
system using terms from the deep learning paradigm. The
study's authors apparently succeeded in developing a mobile
trash pickup robot, as seen in [25]. Using deep learning
techniques, the authors claimed their suggested architecture
could correctly identify trash.

By using a subset of ImageNet in 2012, Krizhevsky et

al. trained one of the biggest CNNs for the ILSVRC-2012
and ILSVRC-2010 competitions [12]. This was the
foundation for the study of Girshick et al. [16]. In what
methods is the detection performance of the Pascal Voc
Dataset evaluated? When RPNs are paired with CNNs, the
resulting approach is called R-CNN. It took a lot of time and
effort to train a regional CNN. It was suggested that SPPnets
[17] be used to accelerate R-CNN. In 2015, Girshick [15]
created Faster R-CNN, a training method that fixes errors Fig 1 Flow Chart of the Presented Method.
produced by R-CNN and SPPnet, making them faster and
more accurate. In 2016, Ren et al. [29] suggested a Faster  Garbage Dataset
Regional CNN to decrease computing time. Keep in mind The largest and highest-quality collection of garbage-
that the convolutional layers used by the Regional Proposal related viral images available for research is found in the
Network detector and the Faster Regional CNN detector are garbage dataset archive [2]. As a whole, the six types of
identical. From what we have learnt about Convolutional garbage included in the collection constitute 2527 images.
Neural Networks (CNNs), we deduced that a faster Regional Figure 2 depicts the range of possible class configurations.
CNN model would be ideal for street rubbish identification. The photographs were arbitrarily assigned to their correct
categories.

IJISRT25SEP396 www.ijisrt.com 610

Volume 10, Issue 9, September – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25sep396

Fig 2 (a) Cardboard, (b) Glass, (c) Metal, (d) Paper, (e) Plastic, and (f) Trash are Classes of the Garbage Dataset.

 Image Pre-Processing  Training, Validation, and Testing

To guarantee more consistent classification results and The garbage dataset produced training, validation, and
better features, each image in the garbage dataset is testing sets. All photo labels were correctly predicted using
preprocessed. Since there is a chance of overfitting while the proposed Garbage Detection using Deep Learning
training using the CNN approach, having a sizable image Method (GD-DLM) on a labelled dataset. The GD-DLM
dataset was essential. model was trained on the training dataset, and its efficacy
was assessed on the validation and test datasets. Therefore,
 Image Resizing we separated our data sets into three equal parts: 60% for
There are 6000 × 4000 variations of each image in the training, 20% for validation, and 20% for testing. Table 1
trash dataset. The size of the dataset has been modified to be shows that 2527 images were used for training, validation,
224 × 224. The model’s performance will be severely and testing on the rubbish dataset. In this work, the model
reduced in exchange for halving the processing time. was trained using data divided into six categories,
representing 60% of all photos. The remaining 40% of
images were used for testing and validation with the rubbish
dataset.

Table 1 Summary of the Garbage Dataset

Split Classes Label Samples Total Samples
cardboard 287
glass 354
metal 286
Training paper 403 1768
plastic 347
trash 91
cardboard 46
glass 65
metal 56
Validation paper 83 328
plastic 61
trash 17
cardboard 70
glass 82
metal 68
Testing paper 108 431
plastic 74
trash 29
Total 2527

IJISRT25SEP396 www.ijisrt.com 611

Volume 10, Issue 9, September – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25sep396

 Architecture of DenseNet121 Base Model Adensely Connected Convolutional Network

The forward pass of a typical Convolutional Neural (DenseNet) describes a network architec ture in which every
Network is quite simple, as seen in Figure 3 below, and it layer communicates with every other layer. L(L+1)/2 direct
results in a predicted label for the input image. Aside from connections between adjacent L-th layers. Each layer feeds
the first convolutional layer, which processes the input image into the next using the feature maps from the layers below it,
directly, subsequent convolutional layers process the output and each layer above it uses the feature maps from the layers
of the preceding layer to generate a feature map. L-direct above it.
connections exist, connecting the previous and next layers in
an L-layered structure.

Fig 3 Architecture of the DenseNet121 Model.

 Architecture of Xception Base Model The Xception neural network design uses depth-
Scientists have developed inception modules that separable convolutions to complete complex tasks. The
bridge the depth-separable convolution method and the engineers at Google came up with the idea. For convolutional
regular convolution utilised in convolutional neural networks neural networks, Google recommends using inception
(a depthwise convolution followed by a pointwise modules as a "middle ground" between regular convolution
convolution). One way to conceptualise a depth separable and the depth-wise separable convolution approach (a
convolution in this context is as an Inception module with a depthwise convolution followed by a pointwise convolution).
maximum height restriction. Using these findings, we With this analogy in mind, it’s easy to see how a depthwise
provide a novel architecture for a deep convolutional neural separable convolution is analogous to an infinitely tall
network that eliminates the need for Inception modules in Inception module. They propose a unique Inception-like
favour of depthwise separable convolutions. On a dataset architecture for deep convolutional neural networks that use
including 350 million pictures and 17,000 classes, our depthwise separable convolutions in place of Inception
Xception [13] approach surpasses Inception V3. However, modules to take advantage of this discovery.
Xception [13] outperforms Inception V3 somewhat on the
ImageNet dataset (for which Inception V3 was built). The Compared to traditional convolutions, depth-separable
Xception design has improved performance through variants are expected to reduce calculation time significantly.
increased capacity and more efficient use of model In Figure 4, we see the analysed data.
parameters.

IJISRT25SEP396 www.ijisrt.com 612

Volume 10, Issue 9, September – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25sep396

Fig 4 Depthwise Separable Convolution

 Evaluation Measures  Recall

The testing dataset was used to evaluate the suggested Another critical parameter is called recall, and it refers
approach following the training phase. We tested the to the percentage of input samples that are of a type that the
architecture's efficacy using recall, accuracy, F1 score, and model can accurately predict. The formula for the recall is as
precision. We will examine the performance metrics utilized follows:
in this research in the parts that follow. What follows is a Recall = T P / (T P+FN)
mathematical definition and representation of the terms "true
positive," "true negative," "false negative," and "false  F1 Score
positive." The F1 score is a statistic utilised to contrast recall and
precision.
 Classification Accuracy
The accuracy of a classification system can be F1 Score = 2 * (Precision * Recall) (Precision + Recall)
evaluated by determining what percentage of its predictions
were correct and what percentage were incorrect. IV. RESULTS AND DISCUSSION

Accuracy = T P+T N / (T P+T N +FP+FN) As part of our training and evaluation, we used high-
powered Graphics Processing Units (GPUs) on a brand-new
 Precision Google Colab [3] Pro account. To accomplish this, we
When analysing the effectiveness of a model, utilized transfer deep learning models. Training of the
classification accuracy may not always be the most proposed GD-DLM with Sparse Categorical Crossentropy
appropriate metric to employ. For instance, this is one of the loss functions was carried out in all trials with the Adam
scenarios where there is a considerable gap in socioeconomic optimizer and a learning rate of 0.0001. During the training
status. It’s a safe bet to assume that each sample is of the phase, which included 10 iterations and an initial batch size
highest possible quality. If the model isn’t picking up any of 8, the best Val_loss models were retained. The Xception
new information, it would be irrational to infer that all and DenseNet121 models recommended the following
components belong to the best class. Therefore, when we talk parameters: 8 batches, 5 epochs, early termination, and
about accuracy, we refer to the fluctuation in findings you model saving based on Val_loss.
receive while measuring the same object several times with
the same tools. The term "precision" refers to one of these  After augmenting it using various methods, we used the
statistics and can be defined as follows: garbage dataset to evaluate the performance of the
provided Xception and DenseNet121 models.
Precision = T P / (T P+FP)  When compared to its predecessors, the suggested GD-
DLM demonstrates significant improvement in terms of
accuracy.

IJISRT25SEP396 www.ijisrt.com 613

Volume 10, Issue 9, September – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25sep396

 The Performance Analysis of the Proposed Garbage be seen in Figure 5, the training accuracy improves from
Detection using Deep Learning Methods (GD-DLM) 72.34% after the first epoch to a final value of 98.53%. As
can be seen in Figure 5, Xception’s validation loss was
 The Performance of Xception Base Model drastically cut down from an initial value of 40.69% to just
The performance of the Xception baseline model was 26.16%. After the first training session, the loss was 85.18%,
measured using the garbage data set. From the end of the first and it was 6.39% after concluding training, which is identical
epoch to the end of the most recent epoch, the accuracy of to the initial loss.
the model validation grew from 86.01% to 92.99%. As can

Fig 5 The Xception Base Model of Accuracy and Loss Graph

Table 2 shows the results of an unobserved test set on had an average F1 score, recall, and precision of 91%, 90%,
the Xception base model. The model’s overall accuracy on and 91%, respectively. The paper’s class performs
the test set was 92.11%; however, Xception performed exceptionally well, with a 95% F1 score, 95% recall, and
exceptionally well on the cardboard class, with a precision of 95% precision. A typical F1 score, precision, and recall for
98%, a recall of 93%, and an F1-score of 96%. The average the plastics category were 89%, 86%, and 92%. For the
values for the F1 score, precision, and recall for the glass rubbish class, the average values for F1 score, precision, and
class were 93%, 92%, and 94%, respectively. The metal class recall were 83%, 92%, and 76%, respectively.

Table 2 Precision, Recall, F1 Score, and Accuracy of the Xception Base Model
Performance Measures Precision Recall F1 Score Accuracy
cardboard 98% 93% 96% 92.85%
Glass 92% 94% 93% 93.90%
Metal 90% 91% 91% 91.18%
Paper 94% 95% 95% 95.37%
Plastic 86% 92% 89% 91.89%
Trash 92% 76% 83% 75.86%
Average Accuracy 92.11%

A confusion matrix was used to assess the classification will be used to evaluate Xception’s performance (shown in
accuracy of different models visually. Rows in the confusion Figure 6). The confusion matrix shows that when the default
matrix that are not on the diagonal represent predictions that parameters for the Xception model are employed, 92.11% of
turned out to be erroneous. Each class’s associated Xception the data are correctly identified, leaving only 7.89%
base model demonstrated that deeper hues suggested more unaccounted for. According to the confusion matrix, the
important classification accuracy, but lighter shades told Xception base model successfully classified all six samples.
misclassified data. The confusion matrix from the test set

IJISRT25SEP396 www.ijisrt.com 614

Volume 10, Issue 9, September – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25sep396

Fig 6 The Xception Base Model Confusion Matrix on Test Set.

 The Performance of DenseNet121Base Model 97.34% after the last epoch. Figure 7 displays a significant
The performance of the DenseNet121 model was reduction in validation loss for DenseNet121, from 54.67%
analysed on the garbage dataset. After the first epoch, the to 40.84%. As an added note, the training loss was 75.03%
model’s validation accuracy was 82.32%; after the most after the first period and 10.11% after the final training, both
recent epoch, it was 88.11%. Figure 6 depicts the training consistent with the initial loss.
accuracy improving from 72.68% after the first epoch to

Fig 7 The DenseNet121 Base Model of Accuracy and Loss Graph.

IJISRT25SEP396 www.ijisrt.com 615

Volume 10, Issue 9, September – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25sep396

Table 3 shows how the DenseNet121 baseline model and recall for the metal class were 89%, 94%, and 91%. The
fared on an unknown test set. While the model’s overall paper class is outstanding, with a 97% F1 score, a 79% recall
accuracy was 86.63% across all classes in the test set, rate, and a 97% precision rate. On average, the F1 score,
DenseNet121 performed best on the cardboard class, with a precision, and recall for the plastics category were 89%,
precision of 79%, a recall of 96%, and an F1-score of 86%. 84%, and 89%. The recall was 87%, the accuracy was 88%,
The average F1 score, precision, and recall for the glass class and the F1 score was 88% on average for the rubbish class.
were 90%, 90%, and 84%. The average F1 score, precision,

Table 3 Precision, Recall, F1 Score, and Accuracy of the DenseNet121 Base Model
Performance Measures Precision Recall F1 Score Accuracy
cardboard 79% 96% 86% 97.71%
Glass 97% 84% 90% 86.25%
Metal 89% 94% 91% 94.12%
Paper 97% 79% 87% 78.70%
Plastic 84% 96% 89% 95.95%
Trash 87% 90% 88% 89.66%
Average Accuracy 88.63%

We could visually compare the models’ ability to shown in the confusion matrix, the DenseNet121 baseline
classify data using a confusion matrix. Diagonal rows in the model’s predictions are consistent across all image types.
confusion matrix represent incorrect predictions. The more The confusion matrix demonstrates that when the
accurate classification was achieved by the corresponding DenseNet121 model was trained with the default settings,
DenseNet121 base model for each class, as indicated by 88.63% of the data were categorised correctly, and 11.37%
darker colors, while lighter colors depicted less precise were misclassified. By comparing confusion matrices across
classification. To evaluate DenseNet121, we will employ all 6 samples, we can see that the DenseNet121 baseline
confusion matrices from the test set (shown in Figure 8). As model achieves excellent results.

Fig 8 The DenseNet121 Base Model Confusion Matrix on Test Set.

Table 4 Classification Accuracy of Proposed Model on Test Set.

Description Model Name Accuracy
GD-DLM Xception Base Model 92.11%
GD-DLM DenseNet121BaseModel 88.63%

IJISRT25SEP396 www.ijisrt.com 616

Volume 10, Issue 9, September – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25sep396

V. CONCLUSION review. Journal of medical Internet research, 2021.

23(5): p. e15708.
The research described in this manuscript explored [10]. Yeo, M., et al., Review of deep learning algorithms
using Convolutional Neural Networks to recognize trauma in for the automatic detection of intracranial
real-time using machine learning techniques. For trauma hemorrhages on computed tomography head imaging.
detection, this method is both reliable and quick. The test Journal of neurointerventional surgery, 2021. 13(4): p.
findings demonstrate a remarkable accuracy rate in 369-378.
identifying people who are either trauma disease or normal. [11]. Hussain, S., et al., Modern diagnostic imaging
The trained model completed the task using the ResNet50V2 technique applications and risk factors in the medical
model, achieving an individual accuracy of 99.40%. field: a review. BioMed research international, 2022.
2022(1): p. 5164970.
The best way to train a CNN model to identify and [12]. Brunner, M., et al., Social media and people with
recognize trauma diseases in humans is to integrate multiple traumatic brain injury: a metasynthesis of research
models and evaluate their performance accuracy. In addition, informing a framework for rehabilitation clinical
the authors recommend an improved optimizer, more precise practice, policy, and training. American journal of
parameter values, enhanced tuning, and models for adaptive speech-language pathology, 2021. 30(1): p. 19-33.
transfer learning. [13]. Buhrmester, V., D. Münch, and M. Arens, Analysis of
explainers of black box deep neural networks for
ACKNOWLEDGMENT computer vision: A survey. Machine Learning and
Knowledge Extraction, 2021. 3(4): p. 966-989.
Special thanks go to Comsats University Islamabad, [14]. Ponnusamy, V., et al., AI‐Driven Information and
Sahiwal Campus, for providing the IT services, Lab, and Communication Technologies, Services, and
server provisions that enabled the smooth writing of the Applications for Next‐Generation Healthcare System.
manuscript. Smart Systems for Industrial Applications, 2022: p. 1-
32.
REFERENCES [15]. Gipson, J., et al., Diagnostic accuracy of a
commercially available deep-learning algorithm in
[1]. Rashid, J., Qaisar, B. S., Faheem, M., Akram, A., supine chest radiographs following trauma. The
Amin, R. U., & Hamid, M. (2024). Mouth and oral British Journal of Radiology, 2022. 95(1134): p.
disease classification using InceptionResNetV2 20210979.
method. Multimedia Tools and Applications, 83(11), [16]. Choi, J., et al., The impact of trauma systems on
33903-33921. patient outcomes. Current problems in surgery, 2021.
[2]. Shifrin, A., et al., Posttraumatic stress disorder 58(1): p. 100849.
treatment preference: Prolonged exposure therapy, [17]. Cheng, C.-T., et al., A scalable physician-level deep
cognitive processing therapy, or medication therapy? learning algorithm detects universal trauma on pelvic
Psychological Services, 2022. radiographs. Nature communications, 2021. 12(1): p.
[3]. Tyagi, A.K. and P. Chahal, Artificial intelligence and 1066.
machine learning algorithms, in Challenges and [18]. Sundkvist, J., et al., Epidemiology, classification,
applications for implementing machine learning in treatment, and mortality of adult femoral neck and
computer vision. 2020, IGI Global. p. 188-219. basicervical fractures: an observational study of
[4]. Ghosh, P., et al., Efficient prediction of cardiovascular 40,049 fractures from the Swedish Fracture Register.
disease using machine learning algorithms with relief Journal of orthopaedic surgery and research, 2021. 16:
and LASSO feature selection techniques. IEEE p. 1-10.
Access, 2021. 9: p. 19304-19326. [19]. Puttagunta, M. and S. Ravi, Medical image analysis
[5]. Ramos-Lima, L.F., et al., The use of machine learning based on deep learning approach. Multimedia tools
techniques in trauma-related disorders: a systematic and applications, 2021. 80(16): p. 24365-24398.
review. Journal of Psychiatric Research, 2020. 121: p. [20]. Tulbure, A.-A., A.-A. Tulbure, and E.-H. Dulf, A
159-172. review on modern defect detection models using
[6]. Bharde, P., R.V. Sai, and S. Sripriya, Role of DCNNs–Deep convolutional neural networks. Journal
computed tomography imaging in the diagnosis of of Advanced Research, 2022. 35: p. 33-48.
blunt and penetrating abdominal trauma injuries. [21]. Gudigar, A., et al., Automated detection and screening
International Journal, 2023. 10(1): p. 10. of traumatic brain injury (TBI) using computed
[7]. Sharma, A.K., et al., Medical image classification tomography images: a comprehensive review and
techniques and analysis using deep learning networks: future perspectives. International journal of
a review. Health informatics: a computational environmental research and public health, 2021.
perspective in healthcare, 2021: p. 233-258. 18(12): p. 6499.
[8]. Chen, Y.-H. and M. Sawan, Trends and challenges of [22]. Pitt, J., Y. Pitt, and J. Lockwich, Clinical and cellular
wearable multimodal technologies for stroke risk aspects of traumatic brain injury, in Handbook of
prediction. Sensors, 2021. 21(2): p. 460. Toxicology of Chemical Warfare Agents. 2020,
[9]. Le Glaz, A., et al., Machine learning and natural Elsevier. p. 745-765.
language processing in mental health: systematic

IJISRT25SEP396 www.ijisrt.com 617

Volume 10, Issue 9, September – 2025 International Journal of Innovative Science and Research Technology
ISSN No: -2456-2165 https://doi.org/10.38124/ijisrt/25sep396

[23]. Qaisar, B. S. (2025). Canker and Cold Sore

Classification Using Xeception Technique. Journal of
Machine Learning and Deep Learning (JMLDL),
2(01).
[24]. Inam Ul Haq; Ahmad Ali; Bilal Shabbir Qaisar; Hafiz
Muhammad Adnan; Mubasher Hussain; Muhammad
Nauman.“Steganography Techniques for Medical
Images: A Recommender Paper.” " Volume. 8 Issue.
9, September - 2023 International Journal of
Innovative Science and Research Technology
(IJISRT), www.ijisrt.com. ISSN - 2456-2165, PP :-
79-87. https://doi.org/10.5281/zenodo.8340603.
[25]. Hamid, S., et al., Dual-energy CT: a paradigm shift in
acute traumatic abdomen. Canadian Association of
Radiologists Journal, 2020. 71(3): p. 371-387.
[26]. Bilal Shabbir Qaisar; Mahammad Ali Shahid; Sunil
Ashraf; Muhammad Adnan; M. Mudasar Azeem;
Maham Ali; Muhammad Nauman. “Automated
Trauma Detection by Using Machine Learning.”
Volume. 10 Issue.8, August-2025 International
Journal of Innovative Science and Research
Technology (IJISRT), 2129-2135
https://doi.org/10.38124/ijisrt/25aug1009.
[27]. Shabbir Qaisar, B., Mudasar Azeem, M., Nauman, M.,
& Yasin, J. (2024). Monkeypox virus detection using
deep learning methods. International Journal of
Computing and Digital Systems, 16(1), 189-198.
[28]. Mudasar Azeem, M., Ul Haq, I., Nauman, M., Talha
Hashmi, M., & Shabbir Qaisar, B. (2024). MASK
DETECTION USING DEEP LEARNING
METHODS. International Journal of Computing and
Digital Systems, 15(1), 1-9.

IJISRT25SEP396 www.ijisrt.com 618