Operations Research and Applications: An International Journal (ORAJ), Vol.12, No.

1, February 2025

LEVERAGING COMPUTER VISION AND NATURAL

LANGUAGE PROCESSING FOR OBJECT DETECTION
AND LOCALIZATION
B. Rahmani 1, S. Bhavanasi 1, A. Maazallahi 2,H. S. Korapala 1, J. L. S. Yenugu 1,
Y.K. Bhatia 1, M.A. Salari 1, E. Snir 3, P. Norouzzadeh, J. Fritts 1
1
Saint Louis University, Saint Louis, USA
2
University of Tehran, Tehra, Iran
3
Washington University in Saint Louis, Saint Louis, USA

ABSTRACT
This paper presents a novel approach leveraging the integration of Computer Vision, Natural
Language Processing (NLP), and Speech Recognition technologies to create an AI-powered
system capable of detecting and locating objects through voice commands. The system developed
using Flask, OpenCV, spaCy, and Blip VQA Model, aims to assist caregivers, visually impaired
individuals, and the elderly in various daily tasks, such as locating items in the home and
checking for potential hazards like leaving the stove on. We also provide the code used for this
project 1.

1. INTRODUCTION
The intersection of computer vision and natural language processing has led to significant
advancements in object detection and localization[1]. Traditional methods often struggle with
detecting concealed, partially hidden, or camouflaged objects, as explored in works like Hidden
and Face-like Object Detection [2], Recognizing Partially Hidden Objects [3][4], and
Camouflaged Object Detection [5]. This study addresses these challenges by developing an
intelligent system capable of detecting and recognizing such objects while offering a user-
friendly interface for natural language-driven object detection queries.

A detection algorithm that fails to ensure high accuracy cannot be applied in real-time scenarios,
especially where safety is crucial[6]. Early methods of object detection relied heavily on
handcrafted features and shallow architectures[7], which limited their applicability. For example,
Late Fusion CNN for Digital Matting[8] highlights the limitations of feature fusion in constrained
environments, while Dense Correspondence [9]introduces more sophisticated correspondence
mapping across scenes. However, both approaches fall short when addressing dynamic and
unpredictable environments. Recent advancements, such as the Deep Learning for Generic Object
Detection Survey [10], demonstrate the shift toward deep learning methods, which offer more
robustness and adaptability in real-time applications. Our approach builds on these advances by
incorporating deep learning architectures capable of understanding both visual inputs and natural
language queries, thereby improving detection accuracy and versatility.

1
https://github.com/b-sai/find_missing_object
DOI : 10.5121/oraj.2025.12101 1
 Operations Research and Applications: An International Journal (ORAJ), Vol.12, No.1, February 2025
The primary objective of this project is to develop an AI-powered system that leverages
advancements in both natural language processing and computer vision to understand user
queries and accurately detect and describe objects within images or live video streams.

Traditional object detection systems such as those based on convolutional neural networks
(CNNs) have demonstrated success in static object detection tasks, but integrating these with
natural language understanding significantly expands their functionality. By incorporating models
like YOLOv3 for real-time detection [11] and Faster R-CNN for speed and accuracy [12], we
enhance our system’s ability to respond dynamically to user queries. This system, designed to be
controlled through voice commands, aims to provide a seamless and accessible experience for
diverse groups, such as caregivers, the visually impaired, and the elderly, by allowing users to
locate objects or detect hazards in real-world settings, like locating misplaced items or identifying
risks like an unattended stove.

He et al.’s Mask R-CNN [13] introduced segmentation alongside object detection, allowing
detailed object recognition. While our focus is on detection and localization, the segmentation
capability provides potential for future enhancements, particularly in applications requiring more
granular object identification.

In the realm of natural language processing, Vaswani et al.’s Transformer model [14] the field by
enabling attention mechanisms, which significantly improve the handling of long-range
dependencies. This framework underpins our use of the BLIP VQA model, which leverages both
visual and language inputs to interpret user queries and deliver accurate object localization.

Finally, Devlin et al.’s BERT model [15] has set a standard for contextual language
understanding, crucial for interpreting user commands in natural language. This contextual
understanding is essential for our system’s ability to accurately process complex or ambiguous
queries, enhancing its utility for a wide range of users.

The integration of vision and language models has become increasingly prominent in recent
years. Radford et al.’s CLIP model [16] introduced a method for training visual models using
natural language supervision, demonstrating robust performance in zero-shot learning and
classification tasks. This vision-language interaction forms the foundation for our system, where
users can locate objects through natural language commands, improving accessibility and ease of
use.

Chen et al.’s work on contrastive learning [17] has also influenced the robustness of visual
representations, focusing on the contrast between positive and negative samples. By
incorporating these principles, our system enhances object detection accuracy, even in complex
environments with occluded or partially visible objects.

Recent advancements in visual question answering, particularly in the work of the BLIP VQA
model, demonstrate the seamless combination of image processing and language understanding.
Our system leverages the capabilities of BLIP VQA to respond to user queries in real-time,
extracting relevant information from both visual input and natural language queries, similar to
applications outlined in Radford et al. [16].

Additionally, survey works such as Zhao et al.’s review of deep learning-based object detection
frameworks [7] provide a comprehensive overview of how convolutional neural networks
(CNNs) have revolutionized the field of object detection. These insights underpin our system’s
architecture, where CNNs play a critical role in ensuring accurate feature extraction and object
localization.

2
 Operations Research and Applications: An International Journal (ORAJ), Vol.12, No.1, February 2025

More recent reviews, such as Amjoud et al. [18], explore the integration of Vision Transformers
with CNNs, presenting a novel approach to real-time object detection. This evolution in object
detection models informs our system’s ability to process visual data efficiently while maintaining
a high level of accuracy, especially in challenging scenarios where traditional CNNs may
struggle.

Recent advancements in object detection using deep learning have also emphasized the
importance of scalability and accuracy in real-world applications. The work by Qiu et al. [19]
provides a comprehensive survey on pre-trained models for NLP, detailing how models like
BERT and GPT have significantly improved language understanding tasks. This is particularly
relevant to our system, as these advancements allow for more nuanced interpretations of user
queries, enabling the system to respond accurately to voice commands for object detection.

Furthermore, Jiao et al. [20] discuss the rapid development of deep learning networks for object
detection, emphasizing their role in security monitoring and autonomous driving. These
applications parallel our system's real-time object detection capabilities, particularly in household
environments were identifying hazards or locating misplaced items is crucial. Our system adapts
these principles to improve accessibility for users such as the visually impaired, enhancing their
independence in daily tasks.

2. METHODOLOGY
The system employs a combination of cutting-edge technologies, including Flask for web
interface development, OpenCV for image processing and object detection, spaCy for natural
language parsing, and the BLIP VQA Model for visual question answering. By integrating these
tools, we enable dynamic interaction where users can transition from static image inputs to live
video streams. This integration significantly enhances user engagement, particularly for
individuals requiring accessibility assistance.

Flask serves as the backbone for the system’s web interface, managing communication between
the user and the backend processes. The lightweight framework ensures low latency, crucial for
real-time object detection. OpenCV, a widely used tool in computer vision, handles the core
image processing tasks. Its real-time capabilities make it ideal for detecting objects in live video
streams and images. OpenCV’s pre-trained models are utilized for initial object detection, which
is then fine-tuned using deep learning models to improve accuracy in detecting occluded or
partially hidden objects.

For natural language processing, spaCy is employed to parse and interpret user commands. The
robust language model enables the system to understand complex queries, allowing for flexibility
in user interaction. By using BLIP VQA, the system further leverages the integration of computer
vision and natural language processing. The BLIP model enables the system to process user
queries by connecting visual data with semantic meaning from the text, ensuring accurate
detection and localization of objects based on user commands.

The architecture of the system incorporates pre-trained models like Faster R-CNN [12] and
YOLOv3 [11], which provide the framework for efficient object detection. The integration of
natural language models, such as BERT [15], enhances the system's ability to understand and
respond to user queries, while the combination of GroundingDINO for object detection enables
zero-shot learning, allowing the system to detect objects that have not been explicitly trained in
the model.

3
 Operations Research and Applications: An International Journal (ORAJ), Vol.12, No.1, February 2025

3. DESIGN ARCHITECTURE
The architecture of the proposed system integrates several core components to facilitate real-time
object detection and localization based on natural language queries. The system leverages Flask
as the web framework, enabling user interaction via a graphical user interface (GUI). This
interface allows users to input natural language queries and receive visual feedback in the form of
detected objects in images or live video streams.

At the core of the image processing pipeline is OpenCV, which performs the necessary pre-
processing and object detection tasks. The input, whether an image or a video frame, is processed
to identify potential objects using traditional image processing techniques alongside deep
learning-based object detection algorithms like YOLOv3 and Faster R-CNN. These models are
selected for their ability to balance accuracy and speed in real-time detection scenarios, ensuring
the system remains responsive and efficient.

To interpret user queries, the system utilizes spaCy, a robust NLP tool that parses the user’s
natural language commands into structured data that the system can process. For example, a
query like “Where is the remote?” is translated into a search for a specific object within the
processed visual data.

The BLIP VQA Model adds a layer of visual question answering, which bridges the gap between
the visual input and the user’s query. The system interprets both the visual features from the
image or video and the linguistic features of the user’s command, making it capable of answering
queries such as “Is there a cup on the table?” or “Locate the stove in this image.”

The entire architecture is designed to facilitate seamless integration between the different
components, ensuring that user commands result in accurate and timely object detection. The
modular nature of the architecture allows for scalability and future updates, including the
potential incorporation of more advanced models or expanded datasets.

Figure 1: This diagram illustrates the architecture of the system, showing the interaction between the client-
side and server-side components. The client-side captures a live video stream, allowing the user to initiate
object detection by submitting a query. The server-side processes the request by analyzing the input image
using a visual question answering (VQA) model and returns a description of the object's location to the
user.

4
 Operations Research and Applications: An International Journal (ORAJ), Vol.12, No.1, February 2025
The architecture of the proposed object detection and localization system consists of two main
components: the client-side and the server-side. On the client-side, a live video stream is
captured, and users can initiate the object detection process by submitting a query, such as asking
for the location of a particular object. This request is then forwarded to the server-side, where the
system utilizes a Visual Question Answering (VQA) model to analyze the input image in
conjunction with the user’s query. The server-side processes the request, detects the object, and
returns a detailed description of the object's location back to the client-side for user feedback.

3.1. Object Detection with GroundingDINO

GroundingDINO[21] represents a significant advancement in object detection, particularly with

its ability to perform zero-shot detection. Traditional object detection models require extensive
labeled datasets for training, but GroundingDINO overcomes this limitation by leveraging vision-
language alignment techniques to detect objects it has not encountered before. This zero-shot
capability is highly valuable in dynamic real-world applications where new objects frequently
appear, and pre-labeling all possible objects is infeasible.

3.1.1. How GroundingDINO Works

GroundingDINO employs a combination of visual feature extraction and textual alignment

techniques. By using a dual-input approach—image and natural language query—the model can
detect objects based on contextual information from both sources. First, GroundingDINO extracts
features from the input image using convolutional neural networks (CNNs) or Vision
Transformers (ViTs). Simultaneously, the model processes the natural language query (e.g.,
“Where is the coffee cup?”) to understand the context of what the user is asking. These two sets
of features—visual and textual—are then aligned through an attention mechanism, allowing the
model to focus on specific parts of the image that are most relevant to the query.

GroundingDINO’s use of attention mechanisms allows it to bypass the need for exhaustive
labeling and training on every possible object. The attention mechanism identifies relevant
regions of the image based on the semantic meaning of the query, directing the model to detect
objects that may not have been explicitly trained. This capability makes it particularly powerful
in zero-shot learning scenarios, where unseen objects can be detected without requiring prior
annotation.

3.2. Zero-Shot Detection in Real-World Applications

The system’s use of GroundingDINO allows it to excel in situations where unknown or novel
objects need to be detected. For example, in household environments, it may be required to detect
a new piece of furniture or an unusual object left out in a hazardous location, such as a knife on a
table. Without requiring new training, GroundingDINO can interpret user queries and identify
these objects in real time, enhancing the system's practicality for safety-critical and everyday
scenarios.

5
 Operations Research and Applications: An International Journal (ORAJ), Vol.12, No.1, February 2025

Figure 2:This diagram represents the architecture for combining visual and language models to generate
meaningful text from image inputs. Queries are processed through a series of pre-trained image and
language models, where the image encoding and query are merged via Q-Former Query Transformers,
allowing the system to generate a language-based output with accompanying visual data.

Once the queries and images are processed, the system employs a Language Learning Model to
generate a natural language description or response, aligning the visual content with the query.
The final step involves the output module, where the system writes a message along with the
photo, producing a meaningful text description or answer related to the image. This architecture
enables enhanced understanding of both visual and language elements, making it suitable for
applications that require interaction between vision and language, such as visual question
answering and complex scene description tasks.

3.3. Experimental Setup

The system was tested in two scenarios:

1. Static Images: A set of 100 images depicting common household objects (e.g., kitchen
appliances, furniture, personal items) was used to evaluate object detection accuracy.
2. Live Video Feeds: The system was also tested in a dynamic environment using live
video feeds to simulate real-time object localization tasks, including identifying partially
hidden objects or camouflaged items in complex visual scenes.

The evaluation was conducted using a combination of pre-trained models (YOLOv3, Faster R-
CNN) for object detection, with the BLIP VQA model and GroundingDINO for visual question
answering and zero-shot detection, respectively. The system's performance was assessed based on
three key metrics: accuracy, response time, and user satisfaction.

3.3.1. Accuracy

The accuracy of the system was measured by its ability to correctly identify and localize objects
based on user queries. In the static image dataset, the system achieved an average accuracy of

6
 Operations Research and Applications: An International Journal (ORAJ), Vol.12, No.1, February 2025
92%, correctly identifying objects in most cases, including occluded and camouflaged items. In
the live video scenario, the accuracy dropped slightly to 88%, mainly due to the dynamic nature
of the environment and occasional misalignment between user queries and object recognition
(e.g., difficulty distinguishing between visually similar objects).

3.3.2. Response Time

Response time is critical in real-time applications, especially when detecting hazards like an
unattended stove. The system's average response time for static image queries was 1.2 seconds,
while in live video feeds, the average response time increased to 2.5 seconds. Although the
response times are adequate for most household tasks, further optimization is necessary to ensure
immediate responses in time-sensitive scenarios (e.g., hazard detection).

3.3.3. User Satisfaction

To evaluate the usability of the system, a group of 20 users, including caregivers and visually
impaired individuals, interacted with the system through voice commands. User satisfaction was
measured through a survey, where participants rated the system on a scale from 1 to 5 based on
its ease of use, accuracy, and usefulness in assisting with daily tasks. The system received an
average rating of 4.3, with users appreciating the ease of interaction through voice commands but
noting occasional delays in object detection during live video streams.

3.4. Example Use Case – Hazard Detection

The system's ability to detect hazards was tested by simulating a scenario where the stove was
left on. The system was able to correctly identify the stove and alert the user through a query like
"Is the stove on?". Using the BLIP VQA model, the system processed both the visual feed and the
query to return the correct result, demonstrating the system's potential for enhancing safety in
household environments.

4. CHALLENGES AND LIMITATIONS

Despite the successful integration of computer vision and natural language processing in the
proposed system, several challenges and limitations were encountered during development and
deployment.

1. Server-Side Processing Constraints:

One of the primary challenges is the computational load on the server, particularly when
processing live video feeds for real-time object detection. The deployment of models like
BLIP VQA and GroundingDINO requires substantial computing power. These models
rely on deep learning architectures that are computationally expensive, limiting the
scalability of the system when deployed on servers with constrained resources. In
scenarios where the system is deployed in real-world applications, such as assisting
visually impaired individuals, the balance between performance and resource constraints
needs to be carefully managed.
2. Resource Constraints for Blip VQA Model:
The BLIP VQA model, which combines visual question answering with object detection,
requires significant computational resources, especially for real-time responses.
Deploying this model on less powerful devices (e.g., mobile platforms or embedded
systems like Raspberry Pi) presents a challenge due to memory limitations and
processing speed. Although this model enhances the system’s versatility in interpreting

7
 Operations Research and Applications: An International Journal (ORAJ), Vol.12, No.1, February 2025
complex user queries, its heavy resource consumption may hinder its use in low-resource
environments.
3. GroundingDINO Deployment:
GroundingDINO's zero-shot object detection capabilities introduce another set of
challenges. While this feature enables the system to detect objects that have not been
explicitly trained on, it requires substantial inference time and computational power.
Efficiently deploying GroundingDINO for real-time applications poses a significant
challenge due to the need for fast query-response times, especially in safety-critical
applications (e.g., identifying hazards in a home environment).
4. NLP Accuracy in Complex Queries:
While the system can handle standard object queries well, there are limitations in
handling more complex queries that involve multiple conditions or ambiguous phrasing.
The spaCy-based natural language processing model can struggle with resolving nuanced
differences in meaning, which may result in less accurate or delayed responses.
Improvements in language parsing, particularly in handling complex grammatical
structures and ambiguous terms, are necessary to fully optimize user interaction.

5. FUTURE DIRECTIONS
While the proposed system demonstrates significant progress in integrating computer vision and
natural language processing for object detection and localization, there are several areas where
further improvements can enhance its capabilities and applicability.

5.1. Hardware Optimization for Edge Devices

One of the main challenges identified is the resource-heavy nature of models like BLIP VQA and
GroundingDINO. Future work could focus on optimizing these models for deployment on low-
power, edge devices such as Raspberry Pi or other embedded systems. This would allow the
system to be more portable and accessible, especially in environments with limited computational
resources, such as smart homes or mobile applications.

5.2. Real-Time Enhancement

Although the system achieves real-time performance in most scenarios, there is room for
improving response times, especially in dynamic environments like live video streams.
Implementing model pruning, quantization, or efficient inference techniques like knowledge
distillation could enhance real-time performance without sacrificing accuracy. Additionally,
exploring asynchronous processing methods could further reduce latency in high-stakes
applications, such as hazard detection.

5.3. Extended NLP Capabilities

While spaCy performs well in processing standard queries, handling more complex or multi-step
user queries is still a limitation. Future research could explore integrating advanced dialogue
systems or multi-turn conversation models to allow the system to understand and process more
complex, nuanced queries. Additionally, integrating multi-lingual NLP capabilities would make
the system more globally accessible, particularly in regions where English is not the primary
language.

8
 Operations Research and Applications: An International Journal (ORAJ), Vol.12, No.1, February 2025
5.4. Improved Object Detection in Complex Environments

GroundingDINO’s zero-shot learning has shown promise, but its performance in highly cluttered
or complex environments needs further testing and improvement. Future iterations of the system
could incorporate attention-based object detection models or transformer-based
architectures like DETR (Detection Transformer) to improve accuracy in detecting and
localizing objects in such challenging conditions.

5.5. Robust Safety Features

Given the system’s potential for real-time hazard detection (e.g., identifying if the stove is left
on), future work could focus on integrating proactive safety features. For example, adding alert
systems that notify users of hazards in real-time, and integrating with smart home systems could
make the system more versatile in preventing accidents or assisting the elderly and disabled in
their daily activities.

5.6. Increased Generalization Across Domains

Currently, the system is tailored for household object detection, but there is significant potential
for expanding its use in other domains such as autonomous driving, healthcare, and industrial
settings. Future research could explore domain-specific adaptations of the system, where fine-
tuning object detection and language models could allow the system to function effectively in
specialized environments.

5.7. Incorporation of User Feedback for Adaptive Learning

To improve the system's long-term effectiveness, integrating user feedback mechanisms could be
a key direction. Allowing the system to learn from user interactions and improve detection
accuracy based on feedback (e.g., correcting false detections) would enable the system to become
more accurate and personalized over time.

6. RESULTS
To evaluate the effectiveness of the proposed system, we conducted a series of experiments
across different real-world environments, focusing on object detection and localization tasks in
household settings. The evaluation aimed to test the system's ability to detect and locate objects
based on user queries, including identifying potential hazards, such as an unattended stove or
misplaced objects like keys and cups.

This study presents an integrated system that leverages computer vision, natural language
processing, and speech recognition to create an AI-powered solution for object detection and
localization. By combining established models like YOLOv3 and Faster R-CNN for object
detection with BLIP VQA and GroundingDINO for visual question answering and zero-shot
learning, the system demonstrates the ability to detect objects in real time and respond to natural
language queries.

The system has shown promising results, particularly in household environments, where it
accurately detects and localizes objects, including concealed and camouflaged items, and
provides descriptions in response to user queries. Its application potential extends to assisting
visually impaired individuals, caregivers, and the elderly, offering a user-friendly interface for
locating objects and identifying hazards. However, the project also highlights certain challenges,

9
 Operations Research and Applications: An International Journal (ORAJ), Vol.12, No.1, February 2025
such as the computational demands of deploying sophisticated models on edge devices and the
limitations in handling more complex natural language queries.

Looking forward to this, there is substantial room for further optimization, including hardware
efficiency, enhanced NLP capabilities, and real-time processing improvements. The system’s
flexibility opens up new possibilities for adaptation in various domains, including autonomous
systems, healthcare, and industrial settings. As the project evolves, continued advancements in
object detection, zero-shot learning, and language understanding will be key to enhancing the
system's functionality and applicability in everyday tasks, ensuring accessibility and safety for all
users.

REFERENCES

[1] S. Chakraborty, “Hidden and Face-like Object Detection using Deep Learning Techniques–An
Empirical Study”.
[2] M. Kowalski, “Hidden object detection and recognition in passive terahertz and mid-wavelength
infrared,” J Infrared Millim Terahertz Waves, vol. 40, no. 11, pp. 1074–1091, 2019.
[3] J. Turney, T. Mudge, and R. Volz, “Recognizing partially hidden objects,” in Proceedings. 1985
IEEE International Conference on Robotics and Automation, 1985, pp. 48–54.
[4] Q. Chen, S. K. Chamoli, P. Yin, X. Wang, and X. Xu, “Imaging of hidden object using passive mode
single pixel imaging with compressive sensing,” Laser Phys Lett, vol. 15, no. 12, p. 126201, 2018.
[5] L. Liu et al., “Deep Learning for Generic Object Detection: A Survey,” Int J Comput Vis, vol. 128,
no. 2, pp. 261–318, 2020, doi: 10.1007/s11263-019-01247-4.
[6] J. Redmon and A. Farhadi, “Yolov3: An incremental improvement,” arXiv preprint
arXiv:1804.02767, 2018.
[7] S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards Real-Time Object Detection with
Region Proposal Networks,” IEEE Trans Pattern Anal Mach Intell, vol. 39, no. 6, 2017, doi:
10.1109/TPAMI.2016.2577031.
[8] K. He, G. Gkioxari, P. Dollár, and R. Girshick, “Mask R-CNN,” IEEE Trans Pattern Anal Mach
Intell, vol. 42, no. 2, 2020, doi: 10.1109/TPAMI.2018.2844175.
[9] A. Vaswani et al., “Attention is all you need,” in Advances in Neural Information Processing
Systems, 2017.
[10] J. Devlin, M. W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional
transformers for language understanding,” in NAACL HLT 2019 - 2019 Conference of the North
American Chapter of the Association for Computational Linguistics: Human Language
Technologies - Proceedings of the Conference, 2019.
[11] A. Radford et al., “Learning Transferable Visual Models from Natural Language Supervision,” in
Proceedings of Machine Learning Research, 2021.
[12] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of
visual representations,” in 37th International Conference on Machine Learning, ICML 2020, 2020.
[13] Z.-Q. Zhao, P. Zheng, S. Xu, and X. Wu, “Object detection with deep learning: A review,” IEEE
Trans Neural Netw Learn Syst, vol. 30, no. 11, pp. 3212–3232, 2019.
[14] A. B. Amjoud and M. Amrouch, “Object Detection Using Deep Learning, CNNs and Vision
Transformers: A Review,” IEEE Access, vol. 11, 2023, doi: 10.1109/ACCESS.2023.3266093.
[15] X. P. Qiu, T. X. Sun, Y. G. Xu, Y. F. Shao, N. Dai, and X. J. Huang, “Pre-trained models for natural
language processing: A survey,” 2020. doi: 10.1007/s11431-020-1647-3.
[16] L. Jiao et al., “A survey of deep learning-based object detection,” IEEE Access, vol. 7, 2019, doi:
10.1109/ACCESS.2019.2939201.
[17] S. Liu et al., “Grounding dino: Marrying dino with grounded pre-training for open-set object
detection,” arXiv preprint arXiv:2303.05499, 2023.
[18] X. P. Qiu, T. X. Sun, Y. G. Xu, Y. F. Shao, N. Dai, and X. J. Huang, “Pre-trained models for natural
language processing: A survey,” 2020. doi: 10.1007/s11431-020-1647-3.
[19] A. B. Amjoud and M. Amrouch, “Object Detection Using Deep Learning, CNNs and Vision
Transformers: A Review,” IEEE Access, vol. 11, 2023, doi: 10.1109/ACCESS.2023.3266093.

10
 Operations Research and Applications: An International Journal (ORAJ), Vol.12, No.1, February 2025
[20] W. Wang, Q. Lai, H. Fu, J. Shen, H. Ling, and R. Yang, “Salient Object Detection in the Deep
Learning Era: An In-Depth Survey,” IEEE Trans Pattern Anal Mach Intell, vol. 44, no. 6, 2022, doi:
10.1109/TPAMI.2021.3051099.
[21] L. Jiao et al., “A survey of deep learning-based object detection,” IEEE Access, vol. 7, 2019, doi:
10.1109/ACCESS.2019.2939201
[22] H. M. Saddique, A. Raza, Z. U. Abideen, and S. N. Khan, “Exploring Deep Learning based Object
Detection Architectures: A Review,” in Proceedings of 2020 17th International Bhurban Conference
on Applied Sciences and Technology, IBCAST 2020, 2020. doi:
10.1109/IBCAST47879.2020.9044558.

11