CYBERNETICS AND SYSTEMS: AN INTERNATIONAL JOURNAL

https://doi.org/10.1080/01969722.2023.2166246

A Deep Learning Approach to Detecting Objects

in Underwater Images
Kalaiarasi Ga, Ashok Ja, Saritha Bb, and Manoj Prabu Mc
a
Department of Electronics and Communication Engineering, V.S.B. Engineering College, Karur,
Tamilnadu, India; bDepartment of Electronics and Communication Engineering, Jai Shriram
Engineering College, Tirupur, Tamilnadu, India; cDepartment of Bio Medical Engineering, Sri
Shakthi Institute of Engineering and Technology, Coimbatore, Tamilnadu, India

ABSTRACT KEYWORDS
A deep learning approach, also known as deep machine learn- Deep learning; underwater
ing or deep structure learning, has recently been found to be pipeline image; object
successful in categorizing digital images and detecting objects detection; aquatic
ecosystem
within them. Consequently, it has rapidly gained attention
and a reputation in computer vision research. Aquatic ecosys-
tems, especially seagrass beds, are increasingly observed using
digital photographs. Automatic detection and classification
now requires deep neural network-based classifiers due to the
increase in image data. The purpose of this paper is to pre-
sent a systematic method for analyzing recent underwater
pipeline imagery using deep learning. There is a logical organ-
ization of the analytical methods based on the recognized
items, as well as an outline of the deep learning architectures
employed. Deep neural network analysis of digital photo-
graphs of the seafloor has a lot of potential for automation,
particularly in the discovery and monitoring of underwater
pipeline images.

1. Introduction
Underwater image recognition is incredibly useful and beneficial for a var-
iety of uses, including pipeline maintenance, mining, marine life monitor-
ing, and military applications. Light moves at a pace of 20 m in clear water
and slightly slower in turbid or coastal water, Veiga et al. (2022).
Underwater light visibility is rather limited, and the speed of entry into the
water drops rapidly. Poor quality underwater photos make it harder to rec-
ognize items. Underwater photography is done at great depths and might
be of poor quality, Mathias et al. (2021). Because underwater photography
is frequently done in deep water, an autonomous underwater vehicle
(AUV) is employed to inspect the photographs.
The artificial light in the AUV intensifies the brightness in the shot, but
it generates a haze and sounds like noise as it moves across the water. To

CONTACT Kalaiarasi G [email protected] Department of Electronics and Communication Engineering,

V.S.B. Engineering College, Karur, Tamilnadu, India.
ß 2023 Taylor & Francis Group, LLC
2 KALAIARASI G ET AL.

increase recognition, underwater images must be preprocessed. The goal of

image preprocessing is to increase image quality by increasing distortion
and visual qualities. Much research has been conducted on this subject, but
it appears to be challenging. Object identification is a computational visual
concept for recognizing things in photos and movies, including identifying
comparable targets in another image, Simonyan et al. (2014). The purpose
of object recognition is to identify items in photographs and detect objects
in the same way that people do. Figure 1 illustrates underwater image
detection.
Images are perceived from several viewpoints, including front, side, and
rear. When an object is partially hidden, viewers experience it in multiple
shapes and sizes, Yan et al. (2013). It recognizes letters, faces, lanes, and
voices, among other things. Water covers more than two-thirds of the
world’s land area, but few tools for studying marine life have been devised.
Marine security, including shipwrecks and naval warfare, is a significant
component of object detection in marine life surveillance, Marini et al.
(2018). The detection of maritime items consists of two steps: feature
extraction and classification. Four geometric features were produced artifi-
cially before recognizing surface vessels, Li et al. (2015). Explicitly learning
sea life can aid in the resolution of maritime challenges such as disaster
avoidance, target detection, emergency rescue, and tracking and detection,
Jalal et al. (2014).
Deep cleaning techniques are utilized in marine systems for data recon-
struction, categorization, and prediction. As part of a deep learning system,
it has two learning layers, a conventional learning layer and a fully conven-
tional learning layer, comprehending complex layers through conversation
and promoting object characterization, Lee et al. (2016). Traditional vector
machines outperform ordinary neural networks and CNNs in object detec-
tion and classification because of their hyperparameters, Yang et al. (2021).
When Constrained Boltzmann Machines or RBM were integrated with

Figure 1. Underwater image detection.

 CYBERNETICS AND SYSTEMS: AN INTERNATIONAL JOURNAL 3

deep learning approaches, learning frameworks grew tremendously. If they

have access to strong GPUs with huge memory capacities in the computer,
it can use the DBN method to supplement deep learning strategy.
When researching deep learning architectures, they merged CNN, RNN,
and AIin the object detection process, yielding good results. A robust train-
ing method with no fake elements enables object detection. Obtain nonlin-
ear information by utilizing CNN capabilities such as weight sharing,
pooling, local connections, and multi-layers. AlexNet’s deep learning suc-
cess in the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC)
is tied to the CNN technique based on VGGNet and ResNet. As a result,
deep CNN findings can provide excellent accuracy as well as data regarding
the deep learning process. DNNs have proven to be successful in the indus-
trial and academic deep learning sectors, according to research. Specific art-
icle recognition in 3 D, as well as visual and object segmentation
recognition, were all processed. Many deep learning and marine life archi-
tectures have been proposed in recent years, but their merits and weak-
nesses remain undefined. As a result, major difficulties in identification
persist. This study’s deep learning mechanisms of marine creatures are
time-based and have both theoretical and practical value in the marine
engineering field.

2. Underwater Object Detection

For many years, the UOD method has been employed in marine ecological
study. Strachan et al. employed color and form descriptors to identify fish
on a conveyor belt monitored by a digital camera. It exhibited a vision sys-
tem for recognizing fish in real-time movies, which included object detec-
tion and tracking techniques. Ravanbakhsh and others To recognize reef
fish, a Histogram of Oriented Gradients (HOG) & Support Vector Machine
(SVM) method was applied. However, the foregoing approaches rely largely
on artisan features, which restricts their ability to express them. With the
rapid evolution of deep learning techniques, a special deep learning-based
UOD technique was recently described. A fast R-CNN-based approach for
identifying and identifying fish species from underwater images is given.In
addition, underwater pictures of plankton items were classified using a
deep residual network model. A lightweight deep neural network for fish
detection based on concatenated ReLU, Inception, and HyperNet is also
reported. In addition, by incorporating multi-scale features and extra con-
textual information, a single-shot feature aggregation deep network for
UOD was presented. The YOLOv2 and YOLOv3 Deep models were refined
and tested on a brackish water dataset that included annotated image
sequences of fish, crabs, and starfish collected in diverse viewpoints.
4 KALAIARASI G ET AL.

Furthermore, UOD’s dataset is quite limited, limiting the development of

his deep learning-based UOD algorithm.

2.1. Deep Learning in Fish Detection and Classification

Prior to 2015, there had been few attempts to apply deep learning to fish
identification. Ravanbakhsh et al. [13] used the haar classifier to classify
shape features. Principal component analysis was used to model the charac-
teristics (PCA). It employs a moving average method to balance the accur-
acy and processing speed of underwater fish recognition. Both systems
have disadvantages when it comes to analyzing big numbers of underwater
photos. Li et al. introduced deep convolutional networks for the first time.
Fish identification and recognition To detect fish, it employed a Fast
Region-based Convolutional Neural Network (Fast-RCNN). We also built a
clean fish dataset of 24272 pictures divided into 12 classes, which is a sub-
set of the ImageCLIEF training and testing datasets. They updated AlexNet
to train Fast R-CNN parameters using stochastic gradient descent (SGD).
Their experimental results demonstrated that larger maximum posterior
estimates resulted in better performance (mAP). These are 9.4% more
accurate on average than the deformable part model (DPM).
In the Fish Knowledge project, Villon et al. evaluated deep learning per-
formance on his GroundTruth dataset. They also compared deep learning’s
performance for fish identification to that of a typical system that used
SVM classification and HOG feature extraction. Their deep network archi-
tecture is modeled by GoogleNet, which has 9 initial layers and 27 layers
with softmax classifiers.

2.2. Deep Learning in Plankton Classification

Plankton are commonly used as markers of ecological health since they
form the foundation of aquatic food webs. For large-scale surveys, trad-
itional plankton monitoring and measurement approaches are insufficient.
In 2015, organized the National Data Science Bowl in partnership with
Oregon State University’s Hatfield Marine Science Center to classify photo-
graphs of plankton. The professor Joni Dambre from the University of
Ghent, Belgium, led the winning team, which used convolutional neural
networks. Deep learning algorithms are widely assumed to require massive
data sets, yet in this case, the classification accuracy is 81.52% with
approximately 30,000 examples in 121 classes.The victorious team’s output
feature map was similar to the input map, pooling and overlapping with a
window size of 3 and an increment of 2. The ultimate structure has 16
 CYBERNETICS AND SYSTEMS: AN INTERNATIONAL JOURNAL 5

layers after starting with a very flat 6-level model and gradually increasing
the amount of layers.
To allow the network to assess the input from numerous perspectives while
using the same feature extraction pipeline, a cyclic pooling method was used.
The same stack of convolutional layers was used, which was then fed into a
stack of dense layers and feature maps that were merged on top.Finally, stacks
of cyclic pooling output feature maps from separate directions are combined
into a single huge stack, and this combined input is used to train the next level,
which includes four times the number of filters as previously. The technique of
integrating feature maps from multiple directions is known as "rolling." It
planned and built an initial module with convolutional layers to avoid distor-
tion and optimize visual information extraction using the same dataset from
the 2015 National Dataset Bowl. GoogleNet served as inspiration. Network
architecture has been defined by increased utilization of computational resour-
ces within a network. To alter rotational and translational invariants, data aug-
mentation was used, and rotational affinity was used to enhance the data.
The deep convolutional neural network was partitioned into two parts:
classifiers and features. However, if the dataset is too tiny, this form of clas-
sifier subdesign overfits, thus it substitute the last two fully connected
layers with small kernels. It turned out to be better than expected. The
model outperformed state-of-the-art approaches at particular image sizes.
Lee and colleagues created a deep network technique for categorizing
plankton using very large datasets. They made advantage of his WHOI
plankton dataset (produced by the Woods Hole Oceanographic Institution).
This dataset included 3.4 million expert-labeled images of him from 103
distinct classes. They primarily concentrated their methods on addressing
the issue of class imbalance in large datasets.
To eliminate the bias induced by class imbalance, they employed the
CIFAR 10 CNN model as a classifier. Three levels of convolutions were fol-
lowed by two completely connected layers in the suggested design. Their
classifier was pre trained using class-normalized data before being
restrained using the original data. We were able to remove the bias from
the class imbalance as a result of this. Dai and colleagues We developed a
deep folding network specifically to classify zooplankton. For the data set,
the ZooScan system acquired 9460 photomicrographs and grayscale pic-
tures of zooplankton from 13 distinct classes. They proposed ZooplanktoN,
a new deep learning architecture for classifying zooplankton. After experi-
menting with various convolution sizes, he determined that ZooplanktoN
net provided the best performance to date at 11 layers.To buttress up their
claims, they ran comparative trials with other deep learning architectures,
including his AlexNet, CaffeNet, VGGNet, and GoogleNet, and discovered
that ZooplanktoN outperformed with 93.7% accuracy.
6 KALAIARASI G ET AL.

2.3. Deep Learning in Coral Classification

Coral color, size, shape, and texture can also differ according to the class.
Furthermore, the border distinctions are hazy and organic. In addition,
currents, algae blooms, and plankton abundance can change water turbidity
and mild availability, impacting image color. Traditional annotation solu-
tions such as bounding boxes, picture labels, and entire segmentation are
ineffective as a result of these issues, Schechner et al. (2005). used the tex-
ture Local Binary Pattern (LBP) and the color Normalized Chromaticity
Coordinate (NCC). They employed a three-layer returned propagation
neural network for categorization. Beijbom et al., on the other hand, were
the first to address automatic annotation on a large scale for coral reef sur-
vey photos by expanding the Moorea Labeled Corals (MLC) collection.
They developed a texture classifier that is completely based on color and
texture descriptors spanning several scales and surpassed existing
approaches. Elawady et al. classified corals by employing supervised
Convolutional Neural Networks (CNNs). They computed Phase
Congruency (PC), Zero Component Analysis (ZCA), and Weber Local
Descriptor using Moorea Labeled Corals and the Atlantic Deep Sea Dataset
from Heriot-Watt University (WLD). They investigated the form and tex-
ture characteristics of entry images while using spatial color channels, Rout
et al. (2018). Mahmood et al. proposed a function extraction technique
based entirely on Spatial Pyramid Pooling to make traditional point-
annotated marine data well matched with CNNs’ entry barriers (SPP). They
used deep capabilities obtained from the VGGNet for coral class, Kumari
et al. (2019). They also combined textual content and color-primarily based
completely handmade features to improve categorization capability.

2.4. Deep Learning Opportunities for Seagrass Detection and Classification

Seagrasses are critical for sediment stabilization, carbon sequestration, and
supplying food and habitat for large marine animals. It is critical to moni-
tor seagrasses in various areas to acquire a better understanding of the tem-
poral and spatial patterns in species composition, reproductive phenology
and frequency, and the consequences of commercialization and human
contact. Ten et al. (2013) used hyperspectral imaging of seagrass ecosystems
to identify tubeworms from the rest of the seagrass surface and performed
binary classification of seagrass. Vasamsetti et al. (2018) undertook a more
extensive analysis to quantify the presence of the seagrass Posidonia oce-
anica in the Gulf of Palma.They made use of analogue RGB data.
The classifier Law’s Logistic Model Tree (LMT) and energy measures
were chosen. A grayscale co-occurrence matrix was used to identify texture
changes. Using sparse coding and morphological filters, Spampinato et al.
(2016) detected seagrass scars on the ocean floor using panchromatic data
 CYBERNETICS AND SYSTEMS: AN INTERNATIONAL JOURNAL 7

taken by orbiting the WorldView2 satellite. This technique was only useful
for detecting coastline scars in flat coastal locations. A common technique
to digital imaging, presently recommended by Australia’s Commonwealth
Scientific and Industrial Research Organization (CSIRO) and Health Safety
and Environment Policies (HSE), is to snap an image from a digital camera
every three seconds.
The camera is normally mounted on a frame and towed behind a boat
cruising at 1.5-3 knots, ensuring that the photo is around 2-3 meters away.
The pictures are then evaluated using the Photo Grid or TranscetMeasure
(VR SeaGIS) programmes. A standard 20-point grid is overlaid, and a human

operator identifies the presence and species of seagrass. Technicians fre-

quently spend several hours analyzing a single transect of 50 m and 25-50
pictures. Because most surveys span hundreds of meters of seafloor, the
analysis can take several days. Furthermore, specialists’ ability to identify
seagrass in pictures varies.
Deep learning algorithms have the potential to increase efficiency while
reducing observer bias from analysis. To the best of our knowledge, there
is no method for detecting seagrass in digital photographs that uses deep
learning. As a result, there is a significant chance to study the deep ocean
floor using deep neural networks to detect and classify seagrass species.

2.5. Methods for Addressing the Issue of Insufficient Underwater Images

Useful training data is required for a deep learning-based UOD model.
Unfortunately, gathering enough picture data in an underwater setting is
difficult. Many solutions have been proposed to address these issues, par-
ticularly data augmentation and image synthesis. Data Augmentation
increases the size of the data collection by changing the labeled data set.
Data can be extended in a variety of ways. Add noise to the image or
reshape it. Recent work has employed generative adversarial neural net-
works to enhance data for domain matching. To generate distorted under-
water views from high-quality aerial RGB or RGB-D shots, image synthesis
methods are given. The image synthesis method is divided into two types:
those based on physical models and those based on deep learning. To
make underwater photographs, physical model-based technologies employ
an underwater image model.
Generative Adversarial Networks have lately been researched in the field
of underwater picture synthesis due to its success in image-to-image
conversion difficulties. It treated underwater image synthesis as an image-
to-image conversion, generating underwater images from RGB-D images col-
lected in the air with a single GAN. To avoid the necessity for image pair
training, it used an unpaired snapshot to train a two-sided cycle consistent
adversarial network to learn interconversions between the air and undersea
8 KALAIARASI G ET AL.

domains. Underwater image synthesis algorithms based on both physical

models and deep learning, on the other hand, are unable to accurately predict
the deterioration history of underwater images, resulting in subpar generated
images.The commonly used underwater physical image model can only syn-
thesis 10 Jerlov water types and takes only two characteristics into account
during the degradation process, leading in considerable flaws in the resultant
image. Additionally, generative adversarial networks are prone to model col-
lapse, which results in images with monotonous tones and frequent artifacts.

3. Proposed Methodology
Object recognition is the ability to properly recognize objects, calculate their
position and dimensions, and conduct semantic or instance segmentation.
Previous studies relied on algorithms based on shape, color, and contour
matching to recognize things, which are unsuitable for real-world object
detection. Deep learning frameworks are categorized as object-regression-
based, classification-based, or both. Methods for recommending regions
include Region-based CNN (RCNN), Fast RCNN, Faster RCNN, and Mask
RCNN, for example, indicate regions of interest and attempt to identify
objects within them, but are incapable of doing so. For direct object detec-
tion, classification-based algorithms make use of an integrated framework.
Figure 2 depicts a flowchart for anticipating underwater images. A convo-
lutional neural network-based underwater object prediction is designed using
deep learning. Using additional number arrays in the model, this network
detects images after encoding them into numeric arrays. Images are recorded
and sent to the internet software as the user adds various characteristics of
the forecast into web form and the model with TensorFlow and scaled it
down from its original enormous size. The numerical values of the various
objects are entered into the data collection by this model.To identify object
prediction from photographs found online, the proposed technique leverages
a CNN object recognition model. Image acquisition, image preprocessing,
segmentation, feature extraction, and grading are five primary stages of object
identification. Scanners are used for image processing tasks such as picture
enhancement, segmenting a photograph into separate sections, locating infec-
tion foci, and extracting features that aid in image classification.

3.1. Dataset
The fundamental data gathering strategy employed in this investigation is
depicted in Figure 3. Kaggle was used to collect data. Image acquisition
refers to the process of capturing images from underwater. A digital camera
or scanner is used to capture images of objects or to collect data.The type
and location of the digital camera have an impact on the quality of the
 CYBERNETICS AND SYSTEMS: AN INTERNATIONAL JOURNAL 9

Figure 2. Flow diagram of object detection in under water images.

images. The initial step in using image data as computational input is to

collect image data.

3.2. Data Preprocessing

As demonstrated in Figure 4, preprocessing methods such as resizing, scal-
ing, zooming, and mirroring were used. Preprocessing occurs after the
image is captured. Enhancing, resizing, enlarging, cropping, changing color
space, smoothing, and eliminating noise from photos are all examples of
preprocessing. Some of the photographs exhibit disorientation, but they
appear to have all been denoised. A distorted image can be enhanced by
removing the distortion with a noise reduction filter. If the image lacks
contrast, actions should be done to improve it.
10 KALAIARASI G ET AL.

Figure 3. Dataset representation.

Figure 4. Preprocessing of input image.

A Gaussian filter was used to soften the image. A histogram graphically

depicts the image intensity distribution. Image contrast with histograms is
improved by anti-aliasing techniques in image processing. This is accom-
plished by dividing the image’s broadest intensity value. Histograms repre-
sent a collection of photos from the preprocessing stage. Each dataset
comprises two image classifications. Datasets for testing and training As a
result, the training and test datasets account for 80% and 20% of the total
dataset, respectively. CNN models are tested on a training set. The valid-
ation set allows to objectively examine the model while fine-tuning the
hyperparameters. The test set analyses the model’s training success by
determining whether the technique used was correct.

3.3. Image Segmentation

After collecting the preprocessed images from the region of interest, image
segmentation is required for object detection. The underwater photographs
in the proposal are broken into portions. Here’s an example of edge detec-
tion in modified segmentation, which separates an image into sections
based on intensity values. In order to detect color changes, it uses the
 CYBERNETICS AND SYSTEMS: AN INTERNATIONAL JOURNAL 11

k-means clustering technique to partition the image and differentiate

between clear and unclear object categories.

3.4. Feature Extraction

Image recognition descriptive features are sought for and extracted as part
of the extraction operation. Color, texture, and shape are all widely
reported characteristics. The primary color components of histograms and
moments are the color qualities that color disease. A texture is created that
depicts the variance of the image texture utilized for disease categorization.
Entropy, uniformity, and contrast are structural characteristics.

3.5. CNN
Because of its ability to perceive and comprehend patterns, CNNs have
evolved substantially. The output accuracy is quite high, making it the
most efficient design for image categorization, retrieval, and recognition
applications (Figure 5). Because it accepts any photo as input, this quality
makes it the best prediction method.A CNN must be able to satisfy crucial
properties such as spatial invariance in order to learn to recognize and
extract visual information from random points within an image. CNNs
automatically extract pictures and learn features from photos and data. As
a result, CNNs can deliver accurate deep learning outcomes.

3.6. Object Prediction

As demonstrated in Figure 2, images are categorized based on the qualities
gathered. Classification is a monitoring technique that divides element pre-
dictions into various categories. Figure 2 depicts how the classifier approach

Figure 5. Convolutional neural network.

12 KALAIARASI G ET AL.

learns to characterize a predefined set of underwater images from a photo-

graph. The training phase is the name given to this stage of learning. The
trained classifier used to assess the images has an impact on accuracy.

4. Simulation Results
The technology recommended will be used in a remotely operated under-
water vehicle (ROV). A total of 33,000 pictures have been manually and
artificially labeled. Deep learning employs 9720 photos for training, 8910
images for validation, and 14370 images for testing. Accuracy, recall, and
mean are often used to assess object detecting accuracy. Using a split train-
ing and testing strategy, divide the data into a 70% training model and a
30% testing model.

 Accuracy: Use the best-fit model to find patterns in the data set.
 Prediction: Predictions with positive outcomes split by total positive
predictions
 Recall: It should be noted that measures are used to establish the type
of TP
 F1 score: Weighted average of precision and recall measurements.

Confusion matrices are used to evaluate and summarize the efficacy of

classifiers.

a. True Positive (TP): Correctly forecast whether the class is positive or

negative.
b. False Positive (FP): An incorrect forecast of a positive class.
c. False Negative (FN): An incorrect categorization prediction.
d. True Negative (TN): Correctly anticipate false classes.

Predictions are used to classify each image in the trained set. The output
of each image is predicted based on its precision.
As illustrated in Figure 6, this technique uses deep learning to produce
results such as object recognition and categorization. So the overall accur-
acy is 98.48%. This can be done with accurate proportions.The training
and validation accuracy are represented on the y-axis while the epoch is
plotted on the x-axis as the model trains the dataset. The link between
training accuracy and epochs is depicted in Figure 7.
The training and validation losses are displayed on the Y-axis and the
epochs are exhibited on the X-axis as the model trains the dataset, as illus-
trated in the graph (Figures 8–10).
 CYBERNETICS AND SYSTEMS: AN INTERNATIONAL JOURNAL 13

Figure 6. Accuracy and loss metrics.

Figure 7. Training accuracy vs epoch.

Figure 8. Validation accuracy vs epoch.

14 KALAIARASI G ET AL.

Figure 9. Training loss vs epoch.

Figure 10. Validation loss vs epoch.

Figure 11. Under water objection detection.

Because more representational information is saved, this approach is

suited for detection in water.In the ROV, a training specific model is uti-
lized to assess the detection effectiveness. The sky is hazy, and the sea is
exceedingly cloudy. Figure 11 depicts the results of the real-time detection.
 CYBERNETICS AND SYSTEMS: AN INTERNATIONAL JOURNAL 15

As illustrated in Figure 11, certain objects are missed because the dataset
is too small, especially if the images in the dataset are so similar. The light-
ing and surroundings are straightforward. As a result, if the trained model
is employed for detection in various sea areas or under different climatic
conditions, the detection accuracy will be diminished. As a result, the pro-
posal intends to capture additional underwater images in diverse marine
locations and climatic conditions.

5. Conclusion
The primary goal of underwater object detection technology is to detect
objects as quickly as possible. It built and tested an autonomous underwater
object detection system that can detect objects in challenging underwater
images in the proposal. The suggested automatic underwater object detection
output is evaluated for accuracy in terms of reduced tracking error compared
to earlier detection approaches. The proposed detection system could be used
as an automated module for an ocean explorer’s high-end computer-equipped
underwater object detection. The proposed method is designed to discover
hidden or camouflaged visualization in underwater situations. The applica-
tion of this method to object detection is widespread, despite its superior
detection accuracy compared to existing systems.

ORCID
Manoj Prabu M http://orcid.org/0000-0002-7316-4810

References
Chiang, J. Y., and Y.-C. Chen. 2012. Underwater image enhancement by wavelength com-
pensation and dehazing. IEEE Transactions on Image Processing : a Publication of the
IEEE Signal Processing Society 21 (4):1756–69. doi:10.1109/TIP.2011.2179666.
Galdran, A., D. Pardo, A. Pic
on, and A. Alvarez-Gila. 2015. Automatic red-channel under-
water image restoration. Journal of Visual Communication and Image Representation 26:
132–45. doi:10.1016/j.jvcir.2014.11.006.
Jalal, A., A. Salman, A. Mian, M. Shortis, and F. Shafait. 2020. Fish detection and species
classification in underwater environments using deep learning with temporal informa-
tion. Ecological Informatics 57:101088. doi:10.1016/j.ecoinf.2020.101088.
Kumari, C. U., D. Samiappan, R. Kumar, and T. Sudhakar. 2019. Fiber optic sensors in ocean
observation: A comprehensive review. Optik 179:351–60. doi:10.1016/j.ijleo.2018.10.186.
Lee, H., M. Park, and J. Kim. 2016. Plankton classification on imbalanced large scale data-
base via convolutional neural networks with transfer learning. In 2016 IEEE
International Conference on Image Processing (ICIP), 3713–7. IEEE.
Li, C.-Y., J.-C. Guo, R.-M. Cong, Y.-W. Pang, and B. Wang. 2016. Underwater image
enhancement by dehazing with minimum information loss and histogram distribution
 16 KALAIARASI G ET AL.

prior. IEEE Transactions on Image Processing : a Publication of the IEEE Signal

Processing Society 25 (12):5664–77. doi:10.1109/TIP.2016.2612882.
Li, X., M. Shang, H. Qin, and L. Chen. 2015. Fast accurate fish detection and recognition of
underwater images with fast r-cnn. In OCEANS 2015-MTS/IEEE Washington, pp. 1–5. IEEE.
Marini, S., E. Fanelli, V. Sbragaglia, E. Azzurro, J. Del Rio Fernandez, and J. Aguzzi. 2018.
Tracking fish abundance by underwater image recognition. Scientific Reports 8 (1):1–12.
doi:10.1038/s41598-018-32089-8.
Mathias, A., and D. Samiappan. 2019. Underwater image restoration based on diffraction
bounded optimization algorithm with dark channel prior. Optik 192:162925. doi:10.1016/
j.ijleo.2019.06.025.
Mathias, A., S. Dhanalakshmi, R. Kumar, and R. Narayanamoorthi. 2021. Underwater object
detection based on bi-dimensional empirical mode decomposition and Gaussian Mixture
Model approach. Ecological Informatics 66:101469. doi:10.1016/j.ecoinf.2021.101469.
Nunes, J. C., S. Guyot, and E. Del
eChelle. 2005. Texture analysis based on local analysis of
the bidimensional empirical mode decomposition. Machine Vision and Applications 16
(3):177–88. doi:10.1007/s00138-004-0170-5.
Rout, D. K., B. N. Subudhi, T. Veerakumar, and S. Chaudhury. 2018. Spatio-contextual
Gaussian mixture model for local change detection in underwater video. Expert Systems
with Applications 97:117–36. doi:10.1016/j.eswa.2017.12.009.
Saleh, A., I. H. Laradji, D. A. Konovalov, M. Bradley, D. Vazquez, and M. Sheaves. 2020. A
realistic fish-habitat dataset to evaluate algorithms for underwater visual analysis.
Scientific Reports 10 (1):1–10. doi:10.1038/s41598-020-71639-x.
Savant, R. R., J. V. Nasriwala, and P. P. Bhatt. 2023. Different skin tone segmentation from
an image using KNN for sign language recognition. In Proceedings of Emerging Trends
and Technologies on Intelligent Systems, 109–18. Singapore: Springer.
Schechner, Y. Y., and N. Karpel. 2005. Recovery of underwater visibility and structure by
polarization analysis. IEEE Journal of Oceanic Engineering 30 (3):570–87. doi:10.1109/
JOE.2005.850871.
Simonyan, K., and A. Zisserman. 2014. Very deep convolutional networks for large-scale
image recognition. arXiv preprint arXiv:1409.1556.
Spampinato, C., S. Palazzo, P.-H. Joalland, S. Paris, H. Glotin, K. Blanc, D. Lingrand, and
F. Precioso. 2016. Fine-grained object recognition in underwater visual data. Multimedia
Tools and Applications 75 (3):1701–20. doi:10.1007/s11042-015-2601-x.
Vasamsetti, S., S. Setia, N. Mittal, H. K. Sardana, and G. Babbar. 2018. Automatic under-
water moving object detection using multi-feature integration framework in complex
backgrounds. IET Computer Vision 12 (6):770–8. doi:10.1049/iet-cvi.2017.0013.
Veiga, R. J., I. E. Ochoa, A. Belackova, L. Bentes, J. P. Silva, J. Semi~ao, and J. M. Rodrigues.
2022. Autonomous Temporal Pseudo-Labeling for Fish Detection. Applied Sciences 12
(12):5910. doi:10.3390/app12125910.
Yan, Z., J. Ma, J. Tian, H. Liu, J. Yu, and Y. Zhang. 2014. A gravity gradient differential
ratio method for underwater object detection. IEEE Geoscience and Remote Sensing
Letters 11 (4):833–7. doi:10.1109/LGRS.2013.2279485.
Yang, H., P. Liu, Y. Hu, and J. Fu. 2021. Research on underwater object recognition based
on YOLOv3. Microsystem Technologies 27 (4):1837–44. doi:10.1007/s00542-019-04694-8.
Yang, H.-Y., P.-Y. Chen, C.-C. Huang, Y.-Z. Zhuang, and Y.-H. Shiau. 2011. Low complex-
ity underwater image enhancement based on dark channel prior. In 2011 Second
International Conference on Innovations in Bio-Inspired Computing and Applications,
17–20. IEEE.

View publication stats