2024 International Conference on Computing and Data Science (ICCDS-2024)

A Comprehensive Review on Underwater Object

Detection Techniques
Prithviraj Guntha P. Mercy Rajaselvi Beaulah
School of Computer Science and Engineering School of Computer Science and Engineering
Vellore Institute of Technology Chennai Vellore Institute of Technology Chennai
Chennai, India Chennai, India
Email: [email protected] Email: [email protected]

Abstract -- Recent years have seen a surge in interest in specializing in detecting minuscule objects. The analysis is
underwater object detection, driven by increased conducted on each approach, considering how they perform
participation in marine research, environmental monitoring, on varying environmental conditions as well as the
2024 International Conference on Computing and Data Science (ICCDS) | 979-8-3503-6533-7/24/$31.00 ©2024 IEEE | DOI: 10.1109/ICCDS60734.2024.10560365

and underwater robotics. This paper offers a systematic computational complexity for these models. Furthermore,
review of recent advancements in this field, covering both
the review encapsulates studies that synergize image
image enhancement and object detection techniques. The
review starts with an examination of state-of-the-art
enhancement with object detection models, emphasizing
underwater image enhancement methods, including selective the intricacies and challenges of optimizing these
color attenuation and feature enhancement modules. Object concurrent tasks. This holistic approach offers insights into
detection approaches are then explored, ranging from potential synergies and areas for refinement in addressing
lightweight neural networks for contaminant detection to the complexities of underwater object detection. In
complex multi-layer models for tiny object features. The conclusion, the study synthesizes overarching trends,
paper provides valuable insights for researchers and identifies persistent challenges, and outlines future
practitioners, summarizing key themes and results while directions in underwater object detection. Considerations
outlining the strengths and limitations of each approach.
such as the efficacy of two-stage versus one-stagedetectors,
Keywords - Image enhancement, object detection, underwater, the impact of network depth, the role of transfer learning,
neural networks, feature enhancement. and the pursuit of lightweight network structures are
meticulously explored. This review aims to serve as a
I. INTRODUCTION technical guide for researchers and practitioners, providing
a road map for navigating the dynamic landscape of
The exploration of the underwater domain is
important for scientific research, environmental monitoring underwater object detection and inspiring future
and autonomous robotics. The fundamental task of innovations in this critical domain.
detecting underwater objects has witnessed a paradigmshift In section 2 of this paper, The various datasets that
owing to advancements in technology and the development are available for building image enhancements and object
of sophisticated methodologies tailored to overcome the detection models are examined. We then move onto section
challenges inherent to submerged environments. This study 3, which talks about the various image enhancement
undertakes a review on cutting- edge techniques and techniques and the various object detection models which
models utilized in underwater object detection, will be discussed in detail. In section 4, we summarise and
encompassing both image enhancement and object review highlighting any key issues that have been found
detection methodologies. The exploration starts withan in- and also discuss improvements and in section 5 we
depth analysis of image enhancement techniques conclude this review.
engineered for peculiarities of underwater conditions.
Techniques such as selective color attenuation, feature II. DATASETS
enhancement modules, and pixel-wise mean squared error
training take the center stage. A discerning examination of Because of the difficulties seen in underwater
these methodologies reveals nuanced insights into their environments, such as limited sight, colour distortion, and
thematic discoveries, achieved results and identified gaps. changing lighting — developing adequate computer vision
models, each with its unique prowess in addressing the models is imperative. These models are crucial for a variety
complexities of underwater scenarios. This study then of activities, including the detection and removal ofmarine
delves into the various object detection Ranging from debris as well as ecological monitoring and species
lightweight neural networks for aquatic contaminant recognition.
detection to competitive multi-layer neural networks

979-8-3503-6533-7/24/$31.00 ©2024 IEEE

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on May 08,2025 at 16:04:59 UTC from IEEE Xplore. Restrictions apply.
 TABLE 1. VARIOUS DATASETS AVAILABLE FOR UNDERWATER OBJECT DETECTION

Dataset Name Number of Classes Number of Images Year of Publication Examples of Objects

F4K 23 27370 2012 Fish species

UIEB 1 950 2019 Underwater images (various scenes)

Brackish 6 14518 2019 Fish, coral

DUO 4 7782 2020 Divers, underwater structures

SUID 1 900 2020 Synthetic underwater objects

TrashCan 22 7212 2020 Trash, ROVs, marine life

UODD 3 3194 2021 Divers, ROVs

CoralNet 7 1741855 2023 Coral reefs, coral species

UWOD 4 8128 2023 Underwater objects (e.g., rocks, plants, animals)

UID2021 52 960 2023 Underwater images (various scenes)

LSUI 18 5004 2023 Underwater scenes (e.g., shipwrecks, reefs,

seagrass meadows)

image quality, making it difficult to discern objects.

The datasets included in the table are important
Enhancing underwater images helps mitigate these
sources of information for practitioners and scholars in the
challenges by improving contrast, reducing noise, and
subject. They give a basis for training and assessing models
revealing crucial details, thereby significantly enhancing
that can function well in difficult aquatic circumstances by
the accuracy and reliability of object detection algorithms
providing a variety of annotated sets of underwater photos.
in underwater scenarios. Once we enhance an image, we
The datasets encompass a broad range of underwater
can then proceed to detect objects from the image. This
situations, such as the identification of fish species,
paper would also investigate the various models
preservation of coral reefs, and detection of various
implemented to detect a wide array of underwater objects
underwater objects.
ranging from small fishes, corals and even garbage present
In table 1, The Fish4Knowledge (F4K) dataset, underwater.
with its focus on fish species, and the UIEB dataset,
In [1], the image enhancement subnetwork
capturing various underwater scenes, contribute to the
enhances object detection by minimizing pixel domainerror
foundational understanding of underwater image analysis.
through a combination of background light and medium
The Brackish dataset, DUO, SUID, and TrashCan datasets
transmission. It employs a feature reconstruction and color
address specific challenges, such as recognizing marine
adjustment block crucial for enhancement. Multi- scale
life, divers, underwater structures, synthetic objects, and
features are eliminated, and dimensions adjusted using a
marine litter, respectively. The datasets introduced in
color adjustment block. Contrast adjustment and MSE loss
recent years, such as CoralNet, UWOD, UID2021, and
ensure brightness and edge sharpness. The RetinaNet
LSUI, showcase the growing emphasis on large- scale,
backbone facilitates object detection, leveraging focused
diverse, and comprehensive collections. These datasets not
classification loss and smooth L1 loss within a multi-term
only advance the state-of-the-art in underwater computer
loss function for accurate localization. The model's dual-
vision but also support broader goals such as coral reef
loss approach enhances spatial coordination and
preservation, marine biodiversity conservation, and
categorization accuracy. An enlarged feature pyramid
understanding underwater ecosystems.
network and dual attention-guided module ensure better
III. IMAGE ENHANCEMENT AND OBJECT feature representation, aiding fine information capture for
DETECTION improved object identification. In [1], while effective, the
model is computationally complex. To address this, [2]
Image enhancement is crucial for underwater introduces YOLOX-Nano, a lightweight neural network
object detection due to the challenging conditions in suitable for resource-constrained environments.
aquatic environments. The scattering and absorption of Enhancements include a Feature Fusion Layer for semantic
light underwater result in poor visibility and degraded and location information utilization, and an Attention

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on May 08,2025 at 16:04:59 UTC from IEEE Xplore. Restrictions apply.
Fusion Module (MYSC) for refining feature channels. Bottom Enhancement augments shallow features
representations. These modifications empower precise of PAFPN with CNN operations, introducing an Attention
detection of aquatic contaminants, particularly addressing module for adaptive weight assignment, improving small
the challenge of recognizing small items underwater, such object detection underwater.
as fish, coral, and microscopic trash. Hence, developing In [8], an enhanced Faster R-CNN method adopts
models capable of accurately detecting small objects a Res2Net101 backbone for superior feature extraction.
underwater is imperative. In [3], TOD-CMLNN excels in GIOU enhances bounding box optimization, while Soft-
identifying small targets by efficiently merging features NMS refines bounding box selection. OHEM mitigates
from different layers through CMLConv and fuzzy positive-negative sample imbalance, and K-means++
learning. Competitive sampling and fusion enhance small clustering ensures efficient box generation. Multi-Scale
item feature description, significantly improving small Training enhances model robustness to diverse object sizes,
object recognition in computer vision. In [4], the UAV making it adaptable to underwater conditions and object
object detection and pose recognition network, based on scales, addressing underwater object detection challenges
ResNet-50, prioritizes computational efficiency and high effectively. In [9], DLRNet, an underwater image
accuracy. The Feature Extraction Module extracts enhancement network, uses a multi-stage process to
multilevel feature maps, refined and fused via the Multi- enhance degraded underwater photos. It steadily improves
scale Feature Fusion Module (BFP). The Detection Head image quality at each level, with a fusion method
tackles foreground-background class imbalance with combining features using an attention module. The three-
VarifocalNet Head and varifocal loss, ensuring reliable stage regression module emphasizes DLRNet's
UAV object identification and posture recognition. effectiveness in advancing underwater image
In [5], AWBiFPN balances accuracy and enhancement, enhancing network understanding and
efficiency by merging feature maps for multi-scale fusion. refining overall enhancement.
A Consistent Supervision Module addresses semantic gaps, In [10], each input batch undergoes dual
while the Residual Feature Augmentation Module augmentation to produce two copies, fed through an
enhances top-level feature representation. The AWBiFPN encoder network resulting in 2048-dimensional normalized
Layer improves accuracy and resource efficiency with embeddings. These embeddings are processed by a
depthwise separable convolutions, effectively detecting projection network during training, omitted during
maritime objects.In [6], proposes a robust solution utilizing inference. Supervised Contrastive loss guides training on
Enhanced Data Augmentation (EDA) and Unsupervised projection network outputs. After training both networks, a
Domain Adaptation (UDA). EDA involves adaptive color linear classifier is trained atop frozen representations using
compensation, CLAHE, and Restorer to enhance cross- entropy loss. This methodology ensures effective
underwater images, addressing color distortion and low representation learning through augmentation, embedding
contrast. UDA handles variations in water quality, generation, and supervised contrastive loss, enabling linear
fostering domain generalization for improved detection classifier training for classification tasks based on frozen
across diverse underwater environments. YOLOv7-Acmix representations. In [11], the FBUD (Feature-Band-Based
integrates the ACmix module to enhance feature extraction Unsupervised Underwater Target Detection) approach
and handle complex underwater scenes. Double Domains aims to identify optimal feature bands for underwater target
Data Augmentation (DDA) augments training set images detection near coasts. Using the unmixing approach and
with original and enhanced images, enhancing accuracy in normalized difference water index (NDWI), it separates
underwater scenes. The SIoU loss function replaces CIoU target and background pixels. FBUD identifies essential
in YOLOv7, incorporating vector angles for faster feature bands through spectral differences between target
convergence and superior positioning accuracy. This and background, enabling simple mathematical procedures
approach combines advanced data augmentation, domain to generate underwater target probability maps. This
adaptation, and loss function modifications, improving methodology highlights the efficacy of feature-based
marine organism detection proficiency in YOLOv7. unsupervised approaches for reliable hyperspectral
In [7], Image Re-annotation corrects object underwater target identification near coasts.
detection dataset errors using Multiple Instance Learning In [12], the CNet network architecture is
(MIL) and symbols for bags, instances, and labels. It introduced for coral seabed image segmentation,
rectifies annotations with a high-accuracy detector, representing a significant advancement in deep learning for
reducing labeling mistakes. The Attention-GB module this domain. CNet features a three-branch parallel encoder
measures channel information using image entropy, structure: RGB encoder (ResNet block), Depth encoder
optimizing attention parameters for red, green, and blue (VGG block), and ShapeConv block-based Fusion encoder.

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on May 08,2025 at 16:04:59 UTC from IEEE Xplore. Restrictions apply.
This architecture enhances segmentation by leveraging of YOLO-based models with Faster R-CNN for underwater
each encoder branch's strengths, making CNet a object identification, providing insight into their structures
noteworthy and comprehensive approach in deep learning- and functions.
based seabed coral reef image segmentation. In [17], the SWIPENet architecture addresses
In [13], TC-YOLO, an extension of YOLOv5s for down sampling difficulties in object detection CNNs.
underwater object detection, integrates Transformer self- Inspired by DSSD, it introduces Hyper Feature Maps with
attention in the backbone and coordinate attention in the high resolution and semantic content, employing dilated
neck. These attention mechanisms enhance feature convolution layers to preserve detailed data. Convolution,
extraction for underwater objects, improving detection dilated convolution, deconvolution, and skip connections
accuracy. Optimal transport label assignment reduces are utilized, along with a unique sample-weighted loss
fuzzy boxes and optimizes training data utilization. This function prioritizing high-weight data for improved feature
innovative approach enhances TC-YOLO's efficacy in learning. SWIPENet's performance in object recognition
underwater object detection, highlighting its notable and noise handling is enhanced. In [18], a neural network
advancement through attention mechanisms and optimal with twelve encoding layers, pre-trained for enhanced
transport label assignment for improved performance. In underwater image processing, is proposed. It employs a
[14], the object identification system employs YOLO for fully convolutional encoder-decoder design with skip
identifying various items and their positions with the slave connections for high-resolution information extraction.
robot's camera. Convolutional neural networks with Initial preprocessing includes underwater image super-
structures containing three convolution layers, two max resolution via ESRGAN. TensorFlow, Adam optimizer,
pooling layers, a fully connected layer, dropout layers, and and data augmentation enhance training. The model aims
a classification layer determine the master robot. This CNN to map natural underwater images to pixel-level semantic
method aims to identify if the spotted item matches the labels in RGB space, prioritizing robustness and superior
target master robot. The model maintains low validation performance in enhancement applications. In [19], the
loss of 0.079 in SCCE with excellent validation accuracy YWnet method enhances YOLOv5 for identifying small
of 99.12%. Additional testing confirms high accuracy of and blurry underwater objects. It integrates SRC3 for
99.59%. However, challenges arise in situations with feature extraction, CBAM to the backbone, and CRFPN to
uneven illumination and contrast. In [15], four foundational the neck for efficient information flow. A decoupled head
processes for underwater target identification in improves accuracy, while EIoU loss enhances performance
hyperspectral imaging are outlined. Background for underwater targets. CRFPN aids multiscale feature
characterization determines optical water characteristics fusion, while CBAM and SRC3 ensure robust feature
using methods like the Case 2 Ecolight model with the extraction. The decoupled head boosts accuracy and
Spectrum Matching Optimization Algorithm (SOMA). convergence speed, with EIoU optimizing regression
Target space modeling computes observed target spectra. accuracy. Phases like data gathering, model training, and
Manifold learning techniques simplify high-dimensional assessment enable real-time underwater target recognition.
hyperspectral data, crucial for effective reduction due to YWnet offers a holistic solution for improving detection in
nonlinear target spaces. Target detection employs challenging underwater conditions.
algorithms on reduced data to distinguish background and The UWV-Yolox model proposed in [20] for
target classes, aiding image target identification. Emphasis underwater video object recognition comprises five main
is on the use of the Spectral Matching Optimization components: input, backbone, neck, prediction, and frame-
Algorithm and manifold learning techniques to enhance level optimization. The input module uses Contrast Limited
detection flexibility. Adaptive Histogram Equalization (CLAHE) to enhance
In [16], the methodology focuses on YOLO and contrast in underwater videos. The backbone incorporates
Faster R-CNN architectures for underwater object the CSP_CA module, integrating Coordinate Attention
detection. YOLO models include versions like YOLOv4 (CA) to improve detection performance in blurry
with improvements like CSPDarknet53 and PAN conditions. The prediction module employs regression and
architecture, optimized for underwater computer vision jitter loss for accuracy, including a novel CIoU loss
with enhancements like Mosaic and CutMix data function. Frame-level optimization involves tubelet
augmentation. Lightweight versions such as YOLOv4 Tiny connecting, re-scoring, and re- coordinating to minimize
are suitable for mobile devices. Modified backbones like detection discrepancies between frames. Tubelet linking
ResNet50-YOLO and DenseNet201-YOLO are explored. joins bounding boxes across frames, while Gaussian
For comparison, Faster R-CNN, a traditional two-stage filtering reduces jitter in coordinate adjustments. This
detector, is used. The paper compares detection outcomes comprehensive approach addresses challenges in

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on May 08,2025 at 16:04:59 UTC from IEEE Xplore. Restrictions apply.
underwater object recognition, yielding reliable results. version exhibit promising results, achieving an
improvement of 1.61% in [email protected]:0.95 without a
IV. EXPERIMENTAL RESULTS noticeable increase in model size and computation
overhead. This emphasizes the significance of models
The review paper provides a comprehensive
tailored for optimal performance in resource-constrained
overview of recent advancements in underwater object
environments. Addressing the challenge of detecting tiny
detection, with a particular focus on image enhancement
objects underwater, TOD-CMLNN [3] stands out with a
techniques and diverse model architectures. The results
dedicated architecture specifically designed for small target
obtained from various models showcase significant
detection, achieving an mAP of 34.10% on the PASCAL
progress in addressing the challenges posed by the unique
VOC2007 test set. The UAV object detection and pose
conditions of underwater environments.
recognition network[4], with its well-structured
In table 2, The table presents a performance architecture, demonstrates balanced and enhanced
comparison of various object detection models based on performance, outperforming Yolov8 by a significant
different metrics. The mean Average Precision (mAP) margin. AWBiFPN[5] proves effective in marine object
reflects the overall precision across object categories, with detection, showcasing a 5.1% mAP increase with ResNet-
higher values indicating superior performance. Frames per 50 compared to BiFPN. However, attention is drawn to
second (FPS) represent the processing speed of each potential information loss during the fusion process and
model, indicating how many images are processed per concerns about the impact of depth-wise separable
second. Average Precision (AP) is a measure of precision convolution on model accuracy. DLRNet[9] introduces a
for individual object categories. Among the models listed, sophisticated multi-stage strategy for underwater image
UAVPNet demonstrates the highest mAP at 88.5%, enhancement, achieving notable improvements in image
showcasing strong overall performance. Despite its slightly details and clarity. The three-stage regression module
lower mAP of 87.4%, YOLOv7-ACmix stands out for its proves optimal, surpassing continuous stage additions. The
remarkably high processing speed, operating at 52 FPS. TC-YOLO[13] network stands out for underwater object
AWBiFPN combines a competitive mAP of 81.63% with a detection, leveraging Transformer self-attention and
processing speed of 12.8 FPS. TOD- CMLNN and optimal transport label assignment to enhance feature
JADSNet exhibit mAP values of 72.46% and 74.41%, extraction, reducing the number of fuzzy boxes and
respectively, with TOD- CMLNN operating at 12 FPS and optimizing training data utilization.
JADSNet at 10 FPS. Improved YOLOX-NANO lags in
However, several challenges and areas for
mAP at 59.1% but maintains a processing speed of 9.5 FPS.
improvement are identified across the models. High
The choice among these models depends on specific use
computational complexity is a common concern, urging the
cases, balancing accuracy, and computational efficiency.
exploration of optimization techniques. Detection of a
TABLE 2. COMPARISON BETWEEN VARIOUS MODELS broader variety of contaminants, addressing haziness,
Model mAP
FPS(frame per
AP
occlusion, and varying lighting conditions, and enhancing
second)
model robustness to diverse environments emerge as
JADSNet 74.41% 10 83 potential avenues for future work. While the models
TOD-CMLNN 72.46% 12 69 demonstrate impressive performance, limitations such as
the need for more effective Underwater Image
AWBiFPN 81.63% 12.8 82
Enhancement (UIE) methods, potential over- reliance on
UAVPNet 88.5% 9.8 90 data augmentation, and limited consideration of water
diversity should be acknowledged. Further research is
Improved
YOLOX- 59.1% 9.5 64 warranted to extend the models' capabilities and ensure
NANO their applicability to real-world underwater scenarios.
YOLOv7-
ACmix
87.4% 52 82 V. CONCLUSION

JADSNet-ResNet101[1] emerges as a top In conclusion, this review paper has provided an

performer with a remarkable mAP of 74.41%, surpassing insightful overview of recent advancements in underwater
RetinaNet-ResNet101 by 2.09%, particularly excelling in object detection, emphasizing image enhancement
small object detection with a mAP of 73.80%. However, it techniques and diverse model architectures. The models
is crucial to note the associated high computational discussed exhibit promising results, showcasing significant
complexity, emphasizing the need for optimization progress in overcoming challenges specific to underwater
strategies or model pruning to mitigate this complexity. environments, such as poor visibility and the detection of
The lightweight YOLOX- Nano[2] and its enhanced small objects. However, a nuanced evaluation reveals

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on May 08,2025 at 16:04:59 UTC from IEEE Xplore. Restrictions apply.
 TC-YOLO with Attention Mechanisms’. Sensors, vol. 23, no. 5,
certain challenges and limitations associated with pp.2567, 2023.
computational complexity, information loss during fusion [14] Lim, D., Kim, J., & Kim, H. 'Efficient robot tracking system using
processes, and the need for more effective Underwater single-image-based object detection and position estimation'. ICT
Image Enhancement (UIE) methods. Looking ahead, future Express, 2023.
[15] Gillis, D. B. ‘An underwater target detection framework for
research should explore optimization strategies, novel hyperspectral imagery’. IEEE Journal of Selected Topics in
fusion mechanisms, and improved UIE methods to address Applied Earth Observations and Remote Sensing. 13, 1798-1810,
these challenges. The development of models with 2020.
enhanced robustness to varying environmental conditions [16] Gašparović, B., Lerga, J., Mauša, G., & Ivašić-Kos,M. ‘Deep
Learning Approach for Objects Detection in Underwater Pipeline
and the detection of a broader range of contaminants would Images’. Applied Artificial Intelligence, 36(1), 2146853, 2022.
further contribute to the practical applicability of [17] Chen, L., Liu, Z., Tong, L., Jiang, Z., Wang, S., Dong, J., & Zhou,
underwater object detection. By focusing on these areas, H. ‘Underwater object detection using Invert Multi-Class
researchers can pave the way for more advanced and Adaboost with deep learning’. International Joint Conference on
Neural Networks (IJCNN) , pp. 1-8, 2020.
applicable solutions, fostering innovations in scientific [18] Adnan Dakhil, R., & Retha Hasoon Khayeat, A. ‘Enhancing
research, environmental monitoring, and autonomous Marine Vision: Deep Learning-Based Underwater Object
underwater robotics. Detection’. International Journal of Computing and Digital
Systems. 15(1), 1- 12, 2024.
REFERENCES [19] Liu, P., Qian, W., & Wang, Y. ‘YWnet: A convolutional block
attention-based fusion deep learning method for complex
[1] Cheng, N., Xie, H., Zhu, X., & Wang, H. 'Joint image underwater small target detection’. Ecological Informatics. 79,
enhancement learning for marine object detection in natural scene'. 102401, 2024.
Engineering Applications of Artificial Intelligence, vol. 120, [20] Pan, H., Lan, J., Wang, H., Li, Y., Zhang, M., Ma, M.,... & Zhao,
105905, 2023. X. ‘UWV-Yolox: A Deep Learning Model for Underwater Video
[2] Wang, Y., Shu, S., Lv, Y., Liu, Y., Zhang, X., & Yan,B. Object Detection’. Sensors, 23(10), 4859, 2023.
'Lightweight Network-Based Aquatic Contaminant Detection
Using a Robotic Vision System'. IEEE Sensors Letters, vol. 7, no.
12, pp. 1-4, 2023.
[3] Chirgaiya, S., & Rajavat, A. 'Tiny object detection model based on
competitive multi-layer neural network (TOD-CMLNN)'.
Intelligent Systems with Applications, vol. 18, 200217, 2023.
[4] Shan, P., Yang, R., Xiao, H., Zhang, L., Liu, Y., Fu, Q., & Zhao,
Y. 'UAVPNet: A balanced and enhanced UAV object detection
and pose recognition network'. Measurement, vol. 222, 113654,
2023.
[5] Gao, J., Geng, X., Zhang, Y., Wang, R., & Shao, K. 'Augmented
weighted bidirectional feature pyramid network for marine object
detection'. Expert Systems with Applications, vol. 237, 121688,
2023.
[6] Zhang, J., Yan, X., Zhou, K., Zhao, B., Zhang, Y., Jiang, H., &
Zhang, J. 'Marine Organism Detection Based on Double Domains
Augmentation and an Improved YOLOv7'. IEEE Access, vol. 7,
pp. 68836-68852, 2023.
[7] Ouyang, W., & Wei, Y. 'An Anchor-free Detector with Channel-
based Prior and Bottom-Enhancement for Underwater Object
Detection'. IEEE Sensors Journal, vol. 23, no. 20, pp. 24800-
24811, 2023.
[8] Wang, H., & Xiao, N. 'Underwater Object Detection Method
Based on Improved Faster RCNN'. Applied Sciences, vol. 13, no.
4, pp. 2746, 2023.
[9] Jia, H., Xiao, Y., Wang, Q., Chen, X., Han, Z., & Tang,Y.
‘Underwater Image Enhancement Network Based on Dual Layers
Regression’. Electronics, vol.13, no. 1,pp. 196, 2024.
[10] Khosla, P., Teterwak, P., Wang, C., Sarna, A., Tian, Y., Isola, P.,&
Krishnan, D. 'Supervised contrastive learning'. Advances in neural
information processing systems, vol. 33, pp. 18661-18673, 2020.
[11] Zhang, S., Duan, P., Kang, X., Mo, Y., & Li, S. ‘Feature Bands
Based Unsupervised Hyperspectral Underwater Target Detection
Near the Coastline’. IEEE Transactions on Geoscience and
Remote Sensing, vol. 61, pp. 5510410, 2023.
[12] Zhang, H., Li, M., Zhong, J., & Qin, J. 'CNet: A Novel Seabed
Coral Reef Image Segmentation Approach Based on Deep
Learning'. In Proceedings of the IEEE/CVF Winter Conference on
Applications of Computer Vision, pp. 767-775, 2024
[13] Liu, K., Peng, L., & Tang, S. ‘Underwater Object Detection Using

Authorized licensed use limited to: Amrita School of Engineering. Downloaded on May 08,2025 at 16:04:59 UTC from IEEE Xplore. Restrictions apply.