See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/371304857

Marine Robotics: An Improved Algorithm for Object Detection Underwater

Article · June 2023

DOI: 10.54105/ijcgm.C7264.082222

CITATIONS READS

0 83

1 author:

Usman Ibrahim Musa

Lovely Professional University
5 PUBLICATIONS   0 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Intracranial-Tumor Detection and Classification System using Convnet and Transfer Learning View project

All content following this page was uploaded by Usman Ibrahim Musa on 05 June 2023.

The user has requested enhancement of the downloaded file.

 Indian Journal of Computer Graphics and Multimedia (IJCGM)
ISSN: 2582-8592 (Online), Volume-2 Issue-2 August 2022

Marine Robotics: An Improved Algorithm for

Object Detection Underwater
Usman Ibrahim Musa, Apash Roy
data from cameras remains a compelling tool for underwater
Abstract: The visibility of items in water is lower than that of
those on land. Light waves from a source don't have enough time
sensing, particularly for close-range detections [1]. The
to reach an item before it vanishes beneath the surface because sensors used by marine robots have different characteristics
light waves in water travel more quickly than they do in air. As a due to the medium they operate in [4]. Detection and tracking
result, it can be challenging for people to deal with water properly experiments have been conducted using marine robots to
due to certain of its physical characteristics. In light of this, object detect artificial targets, and proposed algorithms for color
detection underwater has a wide range of uses, including restoration and detection/tracking of underwater target
environmental monitoring, surveillance, search and rescue, and
navigation. This might enhance the precision, efficiency, and
objects have been demonstrated in experiments with
safety of undersea activities. In light of the aforementioned, this underwater robot platforms [1]. Recently, a method for
paper presents an algorithm for detecting objects underwater detecting seafood items in real-time was suggested that
using YOLOv5. The algorithm has been improved by changing the incorporates the use of Faster R-CNN and the kernelized
way YOLOv5 works, which makes it better at detecting small correlation filter (KCF) tracking algorithm. An underwater
objects. We tested our algorithm and found that it is more accurate picture library was used to train a faster R-CNN detector
than the original YOLOv5 algorithm.
using a VGG model [3]. Small object detection and counting
is an essential need that must be met before underwater
Keywords: Underwater Object Detection, Marine Robotics,
Deep Learning, YOLOv5. robots may be utilised to capture seafood. Marine robots are
essential for completing tasks such as docking station, cable
I. INTRODUCTION tracking, and underwater area coverage for monitoring
purposes [4]. In addition, marine robots are considered an
In recent years, marine robots have gained attention for efficient solution to replace divers in capturing seafood in
marine aquaculture [3]. In research studies using data from
their applications in ocean exploration, marine research, the 2019 China Underwater Robot Competition, the
underwater engineering, and environmental monitoring upgraded YOLOv5s network demonstrated better mean
[1][2]. Underwater object detection has far-reaching Average Precision (mAP) [5]. Due to diverse underwater
significance for the development of marine industry [2]. The surroundings and inadequate training data, existing
development of underwater robots is necessary to achieve the underwater target identification algorithms exhibit
goal of detecting and catching marine small objects [3]. unacceptable accuracy. To increase detection accuracy, a
These robots are at the intersection of the field of robotics and redesigned YOLOv5s network with incorporated CA and SE
oceanic engineering [1]. Underwater robots are useful modules was presented [5].
instruments for identifying items underwater, and their
applications include deep sea infrastructure inspections,
environmental monitoring, and oceanographic mapping [1].
Marine robot performance needs difficult underwater
navigation techniques such as localization, path planning,
and following [4]. Autonomous navigation is critical for
success in executing these activities, especially in underwater
situations where communications are limited [1].
Vision-based object recognition and tracking systems for
underwater robots have been thoroughly investigated. Visual

Manuscript received on 16 August 2022 | Revised Manuscript

received on 18 April 2023 | Manuscript Accepted on 15 May
2023 | Manuscript published on 30 May 2023. Fig. 1 Sample image of how the detection is taking place
*Correspondence Author(s)
Usman Ibrahim Musa*, School of Computer Applications, Lovely
The fig.1 shows a sample image of how the detection is
Professional University, Punjab, India. Email: [email protected],
ORCID ID: https://orcid.org/0000-0002-9659-2340 taking place, where the red boxes indicate objects that have
Apash Roy, School of Computer Applications, Lovely Professional been detected by the algorithm.
University, Punjab, India. Email: [email protected], ORCID ID:
https://orcid.org/0000-0001-6340-6604

© The Authors. Published by Lattice Science Publication (LSP). This is an

open access article under the CC-BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)

Retrieval Number: 100.1/ijcgm.C72640911322 Published By:

DOI: 10.54105/ijcgm.C7264.082222 Lattice Science Publication (LSP)
Journal Website: www.ijcgm.latticescipub.com 1 © Copyright: All rights reserved.
 Marine Robotics: An Improved Algorithm for Object Detection Underwater

Underwater robots are used for various marine pursuits, Using marine robots, object detection can locate occurrences
including object detection. One popular technology for of visual items in digital photos, providing critical
underwater sensing is vision-based object detection, which information for numerous downstream operations [10]. For
works effectively for close range detections [1]. Another example, sea urchins are the primary study focus of aquatic
research approach involves the application of learning-based product detection, and the suggested feature-enhanced sea
visual object detection algorithms to detect debris in the urchin detection method outperforms the traditional
water, such as plastic debris [6]. Convolutional neural Single-Shot Multi Box Detector (SSD) approach [11].
networks (CNN) have been successfully used to detect Moreover, marine robots equipped with underwater object
submerged marine debris with high accuracy, with the detection can improve the accuracy and sensitivity to small
inclusion of different kinds of objects in each class, targets, making them an essential tool for many underwater
potentially improving the variability of object detection [7]. tasks such as robot grabbing tasks of marine products
To identify and detect underwater objects, researchers have [10][11]. So, the advancement of marine robotics has opened
investigated various object detection algorithms and up new avenues for ocean research, and autonomous
state-of-the-art technologies equipped with underwater detection and fishing by underwater robots will be the
robots. For example, a novel object detection algorithm has primary method of obtaining aquatic items in the future.
been proposed for underwater detection, which uses a
A. Impact of Computer Vision and Deep Learning in
combination of deep learning and support vector machine
Underwater
(SVM) techniques to improve its accuracy [2]. Additionally,
researchers have studied the effects of underwater electro Underwater object detection is a challenging task that has
communication and attenuation on the performance of attracted little attention in the realm of lightweight object
different popular object detection algorithms to improve their detection, despite its importance in marine science and repair
effectiveness in underwater environments [8]. Furthermore, and maintenance of underwater structures. The complex
an underwater robot with a length of about 1 m and a width of underwater environment poses a significant challenge for
about 0.8 m has been developed for marine pursuits, detecting underwater targets, as it has low visibility, low
including object detection [9]. Research in this field brings contrast, and color distortion due to illumination [19].
the marine robotics community closer to finding viable Additionally, small underwater objects make detection
methods for detecting and identifying objects underwater. challenging. Traditional object detection algorithms perform
badly in underwater object detection tasks in terms of
accuracy and generalisation [20]. Additionally, underwater
pictures include significant noise, limited visibility, fuzzy
edges, low contrast, colour variation, and crowded
backgrounds, making it a difficult study issue in computer
vision technology [20]. Deep learning has been used to
address a wide range of issues in underwater object detection,
but the benefits and drawbacks of these systems remain
Fig. 2 Object Detection Robot unknown [21]. Light absorption and scattering induced by the
Object detection is a computer vision technology used to medium and suspended particles result in low-contrast and
locate items in images or videos. It entails detecting and haze-like phenomena in underwater photography, resulting in
recognising things of interest within a given picture or video, low-quality movies and photos that make underwater object
such as people, automobiles, buildings, and other objects. To recognition difficult [22][19]. Additionally, the movement of
recognise things in an image or video, object detection image collecting equipment can result in blurring [22].
algorithms often employ a mix of feature extraction, Aquatic organisms often have camouflaged appearances,
classification, and localisation approaches. The use of marine which increase the difficulty of detecting them. The
robots equipped with underwater object detection provides challenges in underwater object detection mainly include
many advantages over traditional methods for exploring the light scattering and absorption, low-quality images and
ocean. These robots play an important part in the operation of videos, multi-scale detection, image enhancement, feature
self-driving marine robots, assisting with operations enhancement algorithms, and accurate detection models that
including course planning, collision avoidance, and control are stable, generalizable, lightweight, and real-time
[10]. Furthermore, marine robots can collect information [20][23][19][20]. An optimal solution for underwater object
about marine animals that is useful for decision-making [10]. detection and species classification does not exist due to these
For example, in commercial fisheries management, marine challenges [23]. The complex and unique nature of
robots may collect critical information for cultivation, status underwater environments presents numerous challenges in
tracking, and disease detection [10]. Furthermore, as object detection, including low visibility, low image quality,
compared to traditional approaches, the utilisation of marine and high variability in appearance. However, computer
robots with underwater object identification gives a vision and deep learning methods have been shown to
significant advantage in ocean exploration [10]. By address these issues effectively [24].
combining marine robots with advanced machine vision
techniques, exploring the underwater environment becomes
much more efficient and effective [10]. Marine robots also
have advantages in object detection compared to traditional
methods for exploring the underwater environment [10].

Retrieval Number: 100.1/ijcgm.C72640911322 Published By:

DOI: 10.54105/ijcgm.C7264.082222 Lattice Science Publication (LSP)
Journal Website: www.ijcgm.latticescipub.com 2 © Copyright: All rights reserved.
 Indian Journal of Computer Graphics and Multimedia (IJCGM)
ISSN: 2582-8592 (Online), Volume-2 Issue-2 August 2022

Deep learning approaches, such as Convolutional Neural object recognition, the SWIPENET+CMA framework
Networks (CNNs), can achieve illumination invariance in provides superior or comparable accuracy in object detection.
underwater object recognition, overcoming the limitations of Although attentional mechanisms have received less
low-quality movies [25]. Because the interest points attention, properly using their potential is likely to effectively
discovered by previous approaches are tiresome and lack improve the development of underwater object detection. In
powerful discriminative information, identifying objects conclusion, while advances in computer vision and deep
based on their forms has been proven to be more successful learning for underwater object recognition have been made
than employing local characteristics [25]. To overcome the recently, there are still numerous obstacles to solve, notably
challenge of finding interesting locations, a deep CNN model in recognising minute items in noisy, low-resolution pictures.
is trained to create abstract discriminative features from Future study should concentrate on overcoming these
low-contrast and low-resolution underwater films [25]. obstacles, including additional examination of
Transfer learning was used to construct a customised CNN attention-based multi-scale feature fusion [27][28].
model for underwater object identification, which overcomes
the restriction of CNNs with limited underwater training data. II. REVIEW CRITERIA
However, due to poor speed and high model size, using
Due to the intricacy of the underwater environment,
DCNNs directly to underwater settings is inefficient [19]. As
detecting underwater objects is a difficult task. The primary
a result, for underwater object identification, a lightweight
obstacles of underwater object identification include poor
detector with high detection accuracy and a small model size
contrast, limited visibility, and the presence of noise. A
is required. Furthermore, the colour conversion module
variety of approaches have been offered to solve these
intends to convert colour photos to grayscale images in order
difficulties. The identification of underwater objects is a vital
to tackle the problem of underwater colour absorption, hence
component of underwater robots and surveillance. For
improving object recognition performance while reducing
underwater object identification, several sensors, such as
computing complexity [26]. Finally, a lightweight deep
acoustic sonar or optical cameras, can be utilised [12].
underwater object identification network has been developed
However, due to strong competitive advantages, vision-based
to address these issues, with promising results in
object identification is favoured [12][13]. Image quality
conventional object detection [26][24]. These new
deterioration, tiny item detection, inadequate generalisation,
approaches have brought solutions to the issues of
and real-time detection are some of the current issues of
underwater object identification, allowing for enhanced
vision-based underwater object detection [13]. Techniques
underwater structural maintenance and repair. Underwater
for detecting underwater objects are mostly developed from
item detection is a difficult task due to interference from the
general object detection and underwater picture enhancement
underwater environment, such as complex backdrop
technologies [14]. Deep learning-based object detection
structures, marine object characteristics, and exploration
algorithms with two and one stages have been employed for
equipment restrictions [27]. Recent advances in computer
underwater item detection. Underwater photos have noise,
vision and deep learning have showed promise in tackling
colour variation, low contrast, blur, and other flaws that make
these issues. Deep learning approaches have sparked a lot of
it difficult to distinguish objects underwater. However,
interest in underwater object recognition because of their
researchers have developed different deep learning-based
ability to directly learn feature representations from data.
underwater object identification systems that have yielded
Underwater object identification based on deep learning
excellent results [14]. Real-time detection in underwater
provides better performance and enormous potential to
object detectors necessitates detection models that are
enhance maritime operations. The review focuses on
accurate, reliable, generalizable, real-time, and lightweight.
vision-based underwater marine item detection, with a
Two major strategies for achieving this objective are
particular emphasis on the detection of marine animals due to
lightweight network architecture and model compression. It
their economic value. The work provides a comprehensive
is advised to iteratively use lightweight network design and
assessment of deep learning-based underwater object
model compression to develop a more elegant model that can
identification approaches and outlines four underwater
be implemented on autonomous underwater vehicles [18].
research challenges: picture quality deterioration, tiny item
While numerous researchers have suggested different deep
detection, inadequate generalisation, and real-time detection
learning-based underwater object recognition approaches,
[26]. In light of the stated issues, a full analysis is offered to
attentional processes for object detection in underwater
provide a clear grasp of the topic [27]. Because of
settings have received little attention. The pipeline that
wavelength-dependent absorption and scattering, underwater
improves photos before object recognition is a popular
photographs frequently contain significant noise, resulting in
strategy in the underwater object detection community, and it
significant visibility loss, contrast reduction, and colour
has been demonstrated to help robots perform underwater
distortion. This noise can mislead detectors and make
missions more effectively [12]. Detecting objects in
identifying tiny objects harder [28]. Deep learning-based
underwater environments is a difficult problem that differs
detectors are currently unsuccessful in detecting small items
greatly from object detection in air.
seen in underwater datasets [28]. Recent research, however,
have offered novel strategies, such as SWIPENET, a
Sample-Weighted hyPEr Network that generates
high-resolution and semantic-rich Hyper Feature Maps,
which improves tiny object recognition [28]. When
compared to many state-of-the-art techniques for underwater

Retrieval Number: 100.1/ijcgm.C72640911322 Published By:

DOI: 10.54105/ijcgm.C7264.082222 Lattice Science Publication (LSP)
Journal Website: www.ijcgm.latticescipub.com 3 © Copyright: All rights reserved.
 Marine Robotics: An Improved Algorithm for Object Detection Underwater

Unbalanced light conditions, poor contrast, occlusion, and tiny object identification include light scattering and
imitation of aquatic animals, which can generate unclear absorption, domain changes, and an imbalance between
objects in photos and videos acquired by underwater positive and negative samples. Overall, underwater object
cameras, bring additional hurdles to object detection [13]. identification is a difficult process that necessitates the use of
Furthermore, generic detectors frequently fail on these specialised equipment and algorithms in order to overcome
ambiguous items, reducing the accuracy of object detection the numerous hurdles offered by the hostile underwater
algorithms. Object identification tasks in aquatic environment.
environments are more difficult than those in terrestrial and
static contexts. As a result, deep learning models for training III. METHODOLOGY
in maritime settings require a huge number of high-quality
Object detection is a process in computer vision where a
photos or videos. The quality of datasets is also critical for computer program can identify and locate objects in a picture
detecting objects in aquatic environments [16]. The water or video. There are different ways of doing this, including
media makes it challenging to get good photos or movies using deep learning or traditional feature-based methods. In
when detecting objects underwater [15]. The viewing this particular study, the focus is on detecting objects that are
distances of any locations match with haze concentration, underwater. We have suggested an improved algorithm
which is a unique constructive cue for salient object called enhanced YOLOv5, which incorporates a new
recognition retrieved from underwater photos. The seeing component called CBAM. This component helps to extract
distance of underwater pictures may be adjusted by the haze valuable features from the image while suppressing features
concentration to provide a unique depth saliency for that are not relevant for object detection. To further improve
underwater landscapes [16]. Despite recent advances in deep the effectiveness of the model, the collar network is replaced
learning, underwater object identification remains a with a BiFPN architecture, and the fusion unit is replaced
difficulty. In underwater object recognition, noisy and with rapid normalized fusion. These changes help to make
inaccurate photos are used as sources of supervision, and the algorithm more efficient in detecting objects underwater.
there are relatively few salient object detection algorithms Overall Network Structure
developed for underwater applications [15][17]. Object
identification technologies for underwater applications are
less effective than those used on land. Furthermore,
underwater object identification algorithms have difficulties
that are not evident in air situations [15].
Due to a variety of circumstances, underwater object
recognition is a substantial difficulty. One of the key
obstacles is a lack of data, as there is no access to huge
datasets that cover a wide range of adverse underwater
circumstances, making it difficult for models to handle the
complicated nature of underwater settings [17]. Furthermore, Fig. 3 Network structure of the improved algorithm
background fluctuation is a significant difficulty in The network architecture of the enhanced technique is
underwater object identification since waves, optical illustrated in Figure 1, which comprises four crucial
diffraction, and different illumination conditions may all components: input, backbone, neck, and detection. This
have an influence on detection accuracy [17]. Most research network retrieves local features from input images through
investigations are limited by the technology employed in various modules such as CBL, CBAMC3, C3, and SPPF. The
underwater object detection, which might result in BiFPN neck structure is equipped with innovative fusion
low-quality photos and movies taken by AUV-mounted modules that blend features of different scales and output the
cameras [18]. When it comes to AUV-based detection, the results via the detecting head. One of the significant
hostile circumstances in undersea/subsea waters complicate components of the network is the convolutional block
object identification, making it a difficult operation. attention module (CBAM) of [17]. It is a flexible and
Furthermore, underwater jobs can be costly and risky for lightweight focus module that can be entirely integrated into
human divers, making high-quality data from these any CNN network. By using CBAM, the model can improve
conditions impossible [18]. Deep learning-based algorithms its performance and interpretability by paying more attention
have challenges due to the low quality of underwater images to the original image. Moreover, CbamC3 is an extension of
and the complexity of underwater settings. Traditional CBAM to C3, as shown in Figure 2. The CBAM module
hand-designed feature extraction algorithms are likewise multiplies the attention map with the input feature map after
unsuitable for genuine underwater scenes, as they are receiving an intermediary feature map, which sequentially
incapable of meeting the requirements for object recognition infers the closely monitoring along two autonomous
in such settings [19]. Furthermore, variances in size or shape, dimensions, i.e., channel and space. The enhanced technique
as well as overlapping or occlusion of marine animals, make network comprises four components that work together to
object detection even more difficult. Furthermore, most retrieve local features from input images and fuse them to
research have an emphasis on low-level feature extraction, detect objects accurately.
which results in poor recognition, low accuracy, and sluggish
recognition [18]. These difficulties make finding an
appropriate solution for underwater item recognition and
species categorization challenging [20]. Other obstacles for

Retrieval Number: 100.1/ijcgm.C72640911322 Published By:

DOI: 10.54105/ijcgm.C7264.082222 Lattice Science Publication (LSP)
Journal Website: www.ijcgm.latticescipub.com 4 © Copyright: All rights reserved.
 Indian Journal of Computer Graphics and Multimedia (IJCGM)
ISSN: 2582-8592 (Online), Volume-2 Issue-2 August 2022

The network architecture includes various modules such as The visual attention module is a component of an artificial
CBAM and CbamC3 that help the model pay more attention intelligence system that helps it focus on specific parts of an
to the original image and improve its performance. image. This module uses a combination of MaxPool and
AvgPool layers to extract important features from the image.
These layers take the output of the previous layer and
combine it to create two feature maps with a single channel
each. These feature maps are then merged together to create a
single feature map with one channel. Next, the spatial
attention feature map is generated. This map helps the system
focus on specific regions of the image that are most relevant
to the task at hand. The spatial attention feature map is
created by passing the merged feature map through a
nonlinear activation function. This function helps highlight
important parts of the image while suppressing irrelevant
details. Once the spatial attention feature map is generated, it
is multiplied by the original image to bring it back to its
original size.

Fig. 4 Network structure of CBAM

The multichannel attention module is shown in the above
illustration. The input map is processed using two Max Pool
and Avg Pool layers simultaneously, and then it is resized to Fig. 6 The spatial attention module
11. Next, the map goes through an MLP module, which To create the CbamC3 module, we combine the CBAM
reduces the number of channels and then increases it back to module with the C3 module. The final output of the
its original number. The Re LU activation function is applied convolutional layer will, as shown in Fig. 5, first pass through
to produce the intermediate output results, which are then the channel attention module in the cbam, then it will pass
combined element by element. A non-linear function is used through with a high dimensional data module, and lastly it
to generate the final output result of the channel attention, will be balanced to achieve the result. To improve extraction
which is multiplied by the original image to produce the final of features, the generated findings are then combined with the
result. The multichannel attention module is a technique used original input utilising the residual structure. To improve the
to enhance the image's features by processing it through backbone network's capacity for extracting features, we
various layers and modules. The input map goes through Max substituted the first and final C3 modules with the CbamC3
Pool and Avg Pool layers and is then resized before passing module after performing preliminary verification.
through an MLP module. The output is then combined and
used to generate the final output result of the channel
attention, which is multiplied with the original image to get
the finished product.

Fig. 7 Network structure of CBAMC3

B. Feature Fusion Network: The weighted bidirectional
feature pyramid network (BiFPN) of [18], a novel feature
matching network that can simply and rapidly execute inter
convolution layer, is used to replace the FPN of YOLOv5. In
order to extract richer feature data from pictures, the BiFPN
structure primarily combines feature maps from the
backbone. The BiFPN structure, as matched to other feature
fusion structures, includes a skip connection that may be used
to access feature data from earlier layers and boost model
effectiveness.

Fig. 5 The channel attention module

Retrieval Number: 100.1/ijcgm.C72640911322 Published By:

DOI: 10.54105/ijcgm.C7264.082222 Lattice Science Publication (LSP)
Journal Website: www.ijcgm.latticescipub.com 5 © Copyright: All rights reserved.
 Marine Robotics: An Improved Algorithm for Object Detection Underwater

In order to fuse the findings of the same layer through a but are not picked up by the network, whereas true positives
cross-scale link, we altered the FPN structure to a BiFPN (TP) signify an object that has been correctly spotted in the
structure, created a connection here between the image. Precision stands for p, recall for r, and category
communications system and the fusion module, and number for N.
illustrated in Fig. 1. This feature fusion will not raise D. Discussions and Results In this experiment, we contrast
computing cost because it is inside the same layer, but it may the original Yolov5s approach with our enhanced version.
boost detection accuracy. We decided on a training batch size of 16 and 150
C. Fusion Module : Features are frequently scaled to the same experimental epochs. Table II displays the final findings.
resolution before being fused when fusing features with
various resolutions. The resolutions of various input Table II. Comparative Results.
characteristics vary, and they make varying contributions to mAP0.95 Recall
the features of the output. Therefore, the weighted fusion YOLOv5 0.5988 0.8753
approach involves giving each input weight and allowing the Ours 0.6062 0.8984
network to determine the significance of each input
characteristic. This procedure, known as Fast normalized
Studies reveal that compared to the original method, our
fusion, is controlled by; When O denotes the weight, I is the
network's learnable weight, ii denotes the input feature map, revised approach has a higher mAP and Recall rate. It shows
and denotes a tiny integer. As a result, we use this novel that our model works better than the YOLOv5.
fusion module, BiFPN Concat, in the neck of YOLOv5, The MAP is raised by roughly 1%, while the Recall is raised
where BiFPN Concat2 and BiFPN Concat3 are fusion by about 2% when compared to the original method.
modules with two and three inputs, respectively. The fusion Fig. 6 displays a visualization of the detection outcomes from
module learns a varied weight for the final fusion based on the WSODD dataset. The detection result of YOLOv5 is on
various inputs, enhancing the network's capacity to fuse the left, while the detection result of our enhanced method is
features and increasing efficiency. The final graph is on the right. As can be observed, the updated algorithm
displayed in Fig. 1. recognizes the target's position with greater accuracy, and the
incidences of missed and false detection have greatly
IV. RESULTS decreased. In summary, our new algorithm performs better
and the identification of tiny targets is enhanced.
To assess the viability and effectiveness of our suggested
approach, we performed tests on the Water Surface Object
Detection Dataset (WSODD). The primary purpose of this
dataset is to solve the issue of limited data on water-based
targets in various settings. The WSODD includes photos
from three different types of water bodies (ocean, lake, and
river) and three weather conditions (sunny, overcast, and
foggy), with a total of 14 different categories of aquatic
objects such as boats and ships. For our experiment, we chose
the WSODD dataset and randomly divided it into training
and validation sets at a ratio of 8:2. The training set contained
5909 photos, while the validation set contained 1558 images.
This allowed us to test our approach on a large dataset with a
variety of aquatic objects and settings. During our
implementation, we used specific hardware and software
indicators, which are listed in Table I. These indicators
played an important role in our experiment, helping us to
ensure that our approach was effective and reliable. By
carefully selecting our hardware and software, we were able
to optimize our approach for the WSODD dataset and
achieve accurate results.
Table I. Fig 8. Visualization of the detection
OS CPU GPU RA FRAMEW VERSIO
M ORK N V. CONCLUSION
Ubuntu Intel NVIDIA 32G Pytorch YOLOv5.
Xeon TESLA T4 6.0 The proposed algorithm was tested on a variety of underwater
Silver environments and the results were promising. The system
4210
uses a combination of image processing algorithms, deep
C. Measures of Evaluation: We use mean average precision learning, and computer vision techniques to accurately detect
(mAP) and recall as the assessment measures to gauge how objects with high accuracy.
well the suggested method works. They come from;
RECALL = TP/TP +FN

Map = AP/N = o1 P(r)dr/N

False negatives (FN) are objects that are present in the picture

Retrieval Number: 100.1/ijcgm.C72640911322 Published By:

DOI: 10.54105/ijcgm.C7264.082222 Lattice Science Publication (LSP)
Journal Website: www.ijcgm.latticescipub.com 6 © Copyright: All rights reserved.
 Indian Journal of Computer Graphics and Multimedia (IJCGM)
ISSN: 2582-8592 (Online), Volume-2 Issue-2 August 2022

The results of this research suggest that object detection in an Marine Object Detection: A Review. Sensors 2023, 23, 1990.
[CrossRef]
underwater environment is feasible and can be used in 11. Kai Hu, Feiyu Lu, Meixia Lu, Zhiliang Deng, Yunping Liu, "A Marine
various applications. Despite the advancements in object Object Detection Algorithm Based on SSD and Feature Enhancement",
detection algorithms for underwater environments, there are Complexity, vol. 2020, Article ID 5476142, 14 pages, 2020. [CrossRef]
still some research gaps that need to be addressed. These 12. Er, M.J.; Chen, J.; Zhang, Y.; Gao, W. Research Challenges, Recent
Advances, and Popular Datasets in Deep Learning-Based Underwater
include the lack of real-world testing and validation, limited Marine Object Detection: A Review. Sensors 2023, 23, 1990.
research on the generalization of algorithms, inadequate [CrossRef]
exploration of environmental factors, insufficient analysis of 13. Er, Meng Joo; Jie, Chen; Zhang, Yani; Gao, Wenxiao (2022): Research
Challenges, Recent Advances and Benchmark Datasets in
computational complexity, and limited research on low-cost Deep-Learning-Based Underwater Marine Object Detection: A Review.
implementations. Addressing these gaps will improve the TechRxiv. Preprint. [CrossRef]
practicality and effectiveness of object detection algorithms 14. Shubo Xu, Minghua Zhang, Wei Song, Haibin Mei, Qi He, Antonio
for underwater environments. Liotta, A systematic review and analysis of deep learning-based
underwater object detection, Neurocomputing, Volume 527, 2023,
Pages 204-232, ISSN
DECLARATION 0925-2312,https://doi.org/10.1016/j.neucom.2023.01.056.
15. Zhe Chen, Hongmin Gao, Zhen Zhang, Helen Zhou, Xun Wang, Yan
Funding/Grants/ Tian, Underwater salient object detection by combining 2D and 3D
No, I did not receive
Financial Support visual features, Neurocomputing, Volume 391, 2020, Pages 249-259,
ISSN 0925-2312,. [CrossRef]
Conflicts of Interest/ No conflict of Interest to the best
16. Ning Wang, Yuanyuan Wang, Meng Joo Er, Review on deep learning
Competing Interests of my knowledge. techniques for marine object recognition: Architectures and algorithms,
No, the Article does not require Control Engineering Practice, Volume 118, 2022, 104458, ISSN
Ethical Approval and 0967-0661,. [CrossRef]
ethical approval and consent to
Consent to Participate 17. Donghwa Lee, Gonyop Kim, Donghoon Kim, Hyun Myung, Hyun-Taek
participate with evidence. Choi, Vision-based object detection and tracking for autonomous
Availability of Data Not Relevant, because it is original navigation of underwater robots, Ocean Engineering, Volume 48, 2012,
and Material/ Data research based on model Pages 59-68, ISSN 0029-8018,. [CrossRef]
18. Dinakaran, R.; Zhang, L.; Li, C.-T.; Bouridane, A.; Jiang, R. Robust and
Access Statement development. Fair Undersea Target Detection with Automated Underwater Vehicles
Usman Ibrahim Musa contributed for Biodiversity Data Collection. Remote Sens. 2022, 14, 3680.
in80% of the work where he [CrossRef]
19. Zhang, M.; Xu, S.; Song, W.; He, Q.; Wei, Q. Lightweight Underwater
worked on the algorithms and the
Object Detection Based on YOLO v4 and Multi-Scale Attentional
Authors Contribution coding part, while Apash Roy Feature Fusion. Remote Sens. 2021, 13, 4706. [CrossRef]
contributed in reviewing the article 20. Sixian Cai, Guocheng Li, Yuan Shan, Underwater object detection using
and suggestions for further collaborative weakly supervision, Computers and Electrical
Engineering, Volume 102, 2022, 108159, ISSN 0045-7906,. [CrossRef]
improvement. 21. Computer Science > Computer Vision and Pattern
Recognition.https://arxiv.org/abs/2209.10151 Accessed 2023-04-17
REFERENCES 22. Chenping Fu, Risheng Liu, Xin Fan, Puyang Chen, Hao Fu, Wanqi
Yuan, Ming Zhu, Zhongxuan Luo, Rethinking general underwater
1. Donghwa Lee, Gonyop Kim, Donghoon Kim, Hyun Myung, Hyun-Taek object dete tion: Datasets, challenges, and solutions, Neurocomputing,
Choi, Vision-based object detection and tracking for autonomous Volume 517, 2023, Pages 243-256, ISSN 0925-2312,. [CrossRef]
navigation of underwater robots, Ocean Engineering, Volume 48, 2012, 23. Fayaz, S., Parah, S.A. & Qureshi, G.J. Underwater object detection:
Pages 59-68, ISSN 0029-8018,. [CrossRef] architectures and algorithms – a comprehensive review. Multimed Tools
2. Sixian Cai, Guocheng Li, Yuan Shan, Underwater object detection using Appl 81, 20871–20916 (2022). [CrossRef]
collaborative weakly supervision, Computers and Electrical 24. L. Chen et al., "Underwater object detection using Invert Multi-Class
Engineering, Volume 102, 2022, 108159, ISSN 0045-7906,. [CrossRef] Adaboost with deep learning," 2020 International Joint Conference on
3. F. Xu et al., "Real-Time Detecting Method of Marine Small Object with Neural Networks (IJCNN), Glasgow, UK, 2020, pp. 1-8, doi:
Underwater Robot Vision," 2018 OCEANS - MTS/IEEE Kobe 10.1109/IJCNN48605.2020.9207506.
Techno-Oceans (OTO), Kobe, Japan, 2018, pp. 1-4, doi: 25. Xin Sun, Junyu Shi, Lipeng Liu, Junyu Dong, Claudia Plant, Xinhua
10.1109/OCEANSKOBE.2018.8558804. Wang, Huiyu Zhou, Transferring deep knowledge for object recognition
4. D. Kim, D. Lee, H. Myung and H. -T. Choi, "Object detection and in Low-quality underwater videos, Neurocomputing, Volume 275,
tracking for autonomous underwater robots using weighted template 2018, Pages 897-908, ISSN 0925-2312,. [CrossRef]
matching," 2012 Oceans - Yeosu, Yeosu, Korea (South), 2012, pp. 1-5, 26. C. -H. Yeh et al., "Lightweight Deep Neural Network for Joint Learning
doi: 10.1109/OCEANS-Yeosu.2012.6263501. of Underwater Object Detection and Color Conversion," in IEEE
5. Wen, G.; Li, S.; Liu, F.; Luo, X.; Er, M.-J.; Mahmud, M.; Wu, T. Transactions on Neural Networks and Learning Systems, vol. 33, no. 11,
YOLOv5s-CA: A Modified YOLOv5s Network with Coordinate pp. 6129-6143, Nov. 2022, doi: 10.1109/TNNLS.2021.3072414.
Attention for Underwater Target Detection. Sensors 2023, 23, 3367. 27. Er, Meng Joo, Jie Chen, Yani Zhang, and Wenxiao Gao. 2023.
[CrossRef] "Research Challenges, Recent Advances, and Popular Datasets in Deep
6. M. Fulton, J. Hong, M. J. Islam and J. Sattar, "Robotic Detection of Learning-Based Underwater Marine Object Detection: A Review"
Marine Litter Using Deep Visual Detection Models," 2019 International Sensors 23, no. 4: 1990. [CrossRef]
Conference on Robotics and Automation (ICRA), Montreal, QC, 28. Long Chen, Feixiang Zhou, Shengke Wang, Junyu Dong, Ning Li,
Canada, 2019, pp. 5752-5758, doi: 10.1109/ICRA.2019.8793975. Haiping Ma, Xin Wang, Huiyu Zhou, SWIPENET: Object detection in
7. M. Valdenegro-Toro, "Submerged marine debris detection with noisy underwater scenes, Pattern Recognition, Volume 132, 2022,
autonomous underwater vehicles," 2016 International Conference on 108926, ISSN 0031-3203,. [CrossRef]
Robotics and Automation for Humanitarian Applications (RAHA),
Amritapuri, India, 2016, pp. 1-7, doi: 10.1109/RAHA.2016.7931907.
8. Yu, H., Li, X., Feng, Y. et al. Multiple attentional path aggregation
network for marine object detection. Appl Intell 53, 2434–2451 (2023).
[CrossRef]
9. Fayaz, S., Parah, S.A. & Qureshi, G.J. Underwater object detection:
architectures and algorithms – a comprehensive review. Multimed Tools
Appl 81, 20871–20916 (2022). [CrossRef]
10. Er, M.J.; Chen, J.; Zhang, Y.; Gao, W. Research Challenges, Recent
Advances, and Popular Datasets in Deep Learning-Based Underwater

Retrieval Number: 100.1/ijcgm.C72640911322 Published By:

DOI: 10.54105/ijcgm.C7264.082222 Lattice Science Publication (LSP)
Journal Website: www.ijcgm.latticescipub.com 7 © Copyright: All rights reserved.
 Marine Robotics: An Improved Algorithm for Object Detection Underwater

AUTHORS PROFILE
Usman Ibrahim Musa is presently enrolled in Lovely
Professional University (L.P.U) in India, where he is
pursuing a bachelor’s degree in computer science. His
areas of research interest primarily encompass machine
learning, computer vision, and cybersecurity. He has
contributed to a few significant projects, including the
development of a real-time skin cancer detection system utilizing machine
learning, the study of cybersecurity and prevention in the quantum era, and
the construction of an intracranial-tumor classification model utilizing CNN.

Apash Roy is currently a full professor at NSHM

Knowledge Campus in Durgapur, India. Prior to this, he
held the position of associate professor at Lovely
Professional University (L.P.U) in India. Professor
Roy's research primarily centers around machine
learning, data science, and pattern recognition. In
addition, he has also worked as an assistant professor at
the University of North Bengal in India.

Disclaimer/Publisher’s Note: The statements, opinions and

data contained in all publications are solely those of the
individual author(s) and contributor(s) and not of the Lattice
Science Publication (LSP)/ journal and/ or the editor(s). The
Lattice Science Publication (LSP)/ journal and/or the
editor(s) disclaim responsibility for any injury to people or
property resulting from any ideas, methods, instructions or
products referred to in the content.

Retrieval Number: 100.1/ijcgm.C72640911322 Published By:

DOI: 10.54105/ijcgm.C7264.082222 Lattice Science Publication (LSP)
Journal Website: www.ijcgm.latticescipub.com
View publication stats
8 © Copyright: All rights reserved.