See discussions, stats, and author profiles for this publication at: [Link]

net/publication/369480723

Deep convolutional neural network with Kalman filter based objected tracking
and detection in underwater communications

Article · March 2023

DOI: 10.1007/s11276-023-03290-z(

CITATIONS READS

2 327

7 authors, including:

Keshetti Srikala Nijil Raj

Mahatma Gandhi Institute of Technology Younus College of Engineering & Technology
9 PUBLICATIONS 43 CITATIONS 6 PUBLICATIONS 31 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Keshetti Srikala on 24 March 2023.

The user has requested enhancement of the downloaded file.

Wireless Networks
[Link] (0123456789().,-volV)(0123456789().,-volV)

Deep convolutional neural network with Kalman filter based objected

tracking and detection in underwater communications
Keshetti Sreekala1 • N. Nijil Raj2 • Sachi Gupta3 • G. Anitha4 • Ashok Kumar Nanda5 • Abhay Chaturvedi6

Accepted: 15 February 2023

 The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2023

Abstract
Underwater autonomous operation is becoming increasingly crucial as a means to escape the hazardous high-pressure
deep-sea environment. As a result, it is essential for there to be underwater exploration. The development of sophisticated
computer vision is the single most significant factor for the success of underwater autonomous operations. In order to
improve low-quality photos and compensate for low-light circumstances, preprocessing is used in underwater vision. This
allows for clearer pictures to be seen. In this paper, we propose a deep convolutional neural network (DCNN) method for
solving the weakly illuminated problem for underwater pictures. This method combines the max-RGB and shade-of-grey
approaches to improve underwater visibility and to train the plotting association necessary to obtain the lighting plot. Using
this method, we are able to resolve the problematic of weakly illuminated pictures in a way that is efficient. After the
photos have been prepared, a deep convolutional neural network (DCNN) approach is developed for detection and
classification in the water. Two updated methods are then utilized in order to adapt the architecture of the DCNN to the
qualities of underwater vision. The purpose of this investigation is to present a Kalman Filter (KF) method as a solution to
the difficulties associated with underwater communication in terms of object tracking and detection. We were able to
separate a section of the object by employing a threshold segment and morphological technique. This allowed us to
investigate the invariant moment and area properties of the section. Based on the findings, it can be decided that the
suggested technique is useful for monitoring underwater targets using DCNN-KF. Furthermore, it displays high resilience,
high accuracy, and real-time characteristics. Results from the simulations show that the suggested model DCNN-KF does a
better job of localization than the most advanced methods at the time of the study.

Keywords Deep convolutional neural networks  Kalman filter  Underwater communications

1 Introduction underwater weapons. In addition, there is a vast array of

potential for the integration of this technology into both the
The technique of underwater target tracking has several military and the civilian sectors. It is able to determine the
applications, including the extraction of marine resources, azimuth and distance of a target by detecting the acoustic
the conduct of underwater engineering tasks, the observa- signals that the target emits, such as the frequency of shell
tion of maritime conflicts, and the homing of accurate vibration or the noise generated by propeller rotation, for

4
& G. Anitha Department of Computer Science and Engineering, Amrita
anitha_g56@[Link] School of Computing, Amrita Vishwa Vidyapeetham,
Chennai, India
1
Department of Computer Science and Engineering, Mahatma 5
Department of CSE, B V Raju Institute of Technology,
Gandhi Institute of Technology, Hyderabad, India Narsapur, Medak, Telangana, India
2
Department of Computer Science and Engineering, Younus 6
Department of Electronics and Communication Engineering,
College of Engineering and Technology, Kollam, Kerala, GLA University, Mathura, U.P., India
India
3
Department of Information Technology, IMS Engineering
College, Ghaziabad, India

123
 Wireless Networks

example. The graphic makes it obvious that each and every deployment of underwater beacons across the working
arrow has an azimuth number between 0 and 360 degrees. zone. In comparison to LBLs, USBLs offer a variety of
The compass bearings can be inducing factor, obscenely benefits, some of which are as follows: the key benefits
high, presumed, recording, or electromagnetic in nature offered by USBL systems are their relatively long range as
depending on the longitude that is being utilized. This well as the ease with which they may be implemented (the
enables it to determine whether or not the target is moving system consists of only two nodes, a transmitter and a
toward or away from the system. The offshore industry and transducer). On the other side, the precision of USBL
mining organizations require underwater survey and measurements diminishes with increasing distance, and
inspection at every stage of the onshore-offshore structure there is also the possibility that multipath difficulties will
installation and operation process. Offshore oil and gas occur. In addition, because of the method in which acoustic
exploration and the mining industry are two of the most waves travel, measurements are sporadic (they occur every
common applications of underwater target tracking today. few seconds), and the passage of time is slowed down in
The first is a survey and examination of the ocean floor, line with the distance between the sending and receiving
followed by IIM (installation, inspection, and maintenance) nodes.
performed deep underwater. This study addresses the sec- In addition to using USBL devices, it is common prac-
ond topic by creating an AUV vision system for inspecting tice to use underwater multibeam sonar systems, which are
and maintaining underwater installations, including also sometimes referred to as acoustic cameras. These
pipelines and cables for oil and gas, as well as electrical systems are used to provide relative position data [3].
power and communications. Underwater structures have Multibeam sonars used today produce an acoustic image
become increasingly popular. To keep them safe from the with a high frequency and great precision; nevertheless,
hazards of fishing and anchoring, regular examination and their field of view is limited and their range is limited as
maintenance is recommended [1]. The presence of noise on well. In contrast to USBLs, sonars need additional acoustic
a subsea surface is rather common, and as a result, locating image processing in order to detect where an object is
and following the path of an underwater pipeline in a located inside the field of view. To find submerged items as
maritime environment that is otherwise complex is an well as gauge the depths of the ocean, SONAR (sound
endeavor that is fraught with difficulty. Most of the noise in tracking and reaching) is employed. Because of the pres-
underwater photos comes from the uneven growth of ence of background noise, this processing frequently
marine life and the way the light changes. results in an inaccuracy being introduced into the mea-
The inability to obtain global positioning signals, which surement. Armed services acoustics, offshore drilling, and
are normally accessible in locations that can receive commercial transportation produce the harshest and most
satellite transmissions, presents a significant challenge for obtrusive manmade coastal waves, as well as the Center is
the field of marine robotics when it comes to the tracking of fighting to preserve endangered sea creatures from every
targets. Underwater localization and navigation systems one of these dangers.
that make use of acoustic-based sensors, such as LBL The active sonar echo signal contains information on the
(long-baseline), SBL (short-baseline), and USBL (ultra- scattering characteristics of the target that can be utilized to
short-baseline), triangulate responses received from audio locate and identify the target. The term sonar can be used to
beacons [2]. One of the three primary types of underwater describe one of two types of equipment passive sonar
acoustic location technologies utilized to monitor explorers involves paying attention for the audio created by water-
as well as submarines is the long baseline (LBL) acoustical craft, while active sonar entails emitted sound impulses and
global positioning. By detecting the user’s distances from returns detection. Sonar can be employed to locate objects
three or more sensors, that are, for instance, dropped over in the ocean by sound as well as to analyze their echo
the outside of the small boat out of which surveillance patterns. In the discipline of underwater acoustics, scien-
activities are conducted, short benchmark devices can tists have utilized board sprinkling properties for goal
establish the location of a monitored object, like a ROV. discovery or categorization. A radar antenna is a device
Ultra-short baseline, commonly referred to as SSBL for that sends out electromagnetic radiation and listens for
Super Short Base Line, is a technique for placing an their reflections. The capacity of a transmitter to identify
underwater environment system. It is employed to follow the path to follow that an item is situated determines how
underwater objects like ROVs, AUVs, or explorers. This is well it performs. Gaussian classification in machine
done in order to get around the challenge presented by the learning categorized by extracting structures from the goal
aforementioned difficulties. USBLs, which allow relative echo sign in a way that mimics auditory perception [4, 5].
underwater localization via sonic propagation, are gener- When it comes to multi-target detection and classification,
ally utilised for tracking underwater objects. This is in deep learning shines because of its ability to automatically
contrast to LBLs, which ask for the bothersome extract target feature information from raw data via

123
Wireless Networks

training. Computer vision, NLP, speech recognition, and look into the DCNN-KF modelling approaches that can be
object detection are just a few of the domains where the used for modelling underwater pipelines and shape space
widespread use of convolutional neural networks (CNNs) transformation. In Unit 4, we will deliberate the conse-
has led to significant productivity and efficiency gains quences of testing the proposed system on actual under-
[6–8]. Convolutional neural networks also known as CNNs water photographs, and in Sect. 5, we will discuss the
is used in conjunction with previously developed models of prospects for the suggested model moving forward.
auditory perception in order to replicate the operation of an
entire auditory system in instruction to classify ship targets
from their radiated noise [9]. A statistic, or functional of 2 Literature survey
the information, called as a assessor or terminal estimation
is employed to figure out the significance of an unidentified In addition to this, the Kalman filtering approach is used
variable in a predictive method. The expression the esti- extensively in the industry of underwater target tracking.
mator is the approach employed to generate an approxi- Inside the matrices of the filtration system, screening,
mation of an unknown variable is a popular one. This was desorption, and entrapment are the three main filtering
accomplished in instruction to classify ship targets from processes. In order to send air and certain other vapors to
their radiated noise. On a radar scope placed on a space- antiseptic regions, sterilising graded filtration are employed
ship, a variety of misleading response back, such as to handle heat-sensitive needles, ophthalmology treat-
interfering reverberations, secondary tracing resounds, or ments, biologicals, or atmosphere. The hybrid unscented
passband resonates, can be seen. The findings point to the Kalman filter was the name that Kumar [11] gave to his
viability of applying deep learning to the subsea audibility novel estimator that he suggested. This novel estimator
industry. A deep learning approach is planned to build a takes three different approaches that are already in use and
model for extracting classification features from underwa- combines them to create a result that is substantially more
ter acoustic signals using a generative confrontation net- accurate than any of those methods individually. The
work and a deep neural network classifier for inflection methods that are combined are the UKF, the integration
acknowledgement, and demonstrated the model’s ability to methodology, and the preprocessing device (see below).
extract cataloguing features from such signals [10]. He did Reference [12] used an enhanced colour Kalman filter to
all of this work using a generative confrontation network predict the precise prototypical and insolence of transmis-
and a deep neural network classifier. The use of broad deep sion line examination UAVs. This ensured that the UAVs
learning in the context of underwater acoustic targets pre- would maintain a steady, dependable, and safe flying despite
sents a number of exciting opportunities. the presence of electromagnetic interference. On the other
In this study, we suggest using a deep convolutional hand, the methods that were just discussed have a few
neural network (DCNN) to train a mapping relationship to drawbacks. The conventional Kalman filter is an efficient
create an illumination map for underwater photos, which estimate method, but it only works with linearization. Since
improves visibility by combining the max-RGB and shade- typical implementations are complex, Kalman filters aren’t
of-grey approaches. Using the results of image processing, really applicable. The Kalman filtering approach makes use
a deep convolutional neural network (DCNN) strategy is of information that is more accurate from the past, and the
created for performing detection and organization in the fact that the actual sonar system may be subject to a variety of
ocean, and two updated methods are utilised to modify the unknown disturbances may result in significant variations in
DCNN structure according to the characteristics of under- estimations. With a common wavelength of 1.5 m–1.5 cm,
water vision. This study proposes a Kalman Filter (KF) acoustic sensors employed in combat must be capable of
method to address the challenges of underwater commu- detecting objects at an acceptable distance, which limits the
nication regarding object tracking and detection. Properties operating resonant frequency to roughly 1–100 kHz.
of invariant moment and area were analysed after a region In order to solve classification issues, Krizhevsky et al.
of interest was identified from the object using a threshold [13] developed a CNN technique, which allowed them to
segment and morphological approach. According to the become the champion of the ILSVRC (ImageNet Large
findings, the method that is provided for tracking under- Scale Visual Recognition Challenge) and brought the top 5
water targets using DCNN-KF is effective, and it has a fault rate down to 15.3%; ever since then, deep CNN has
number of desirable characteristics such as high resilience, been widely utilized. The RPN (Region Proposal Network)
high precision, and real-time tracking. and CNN algorithms have both been validated on the
The rest of the article is structured as follows: in Sect. 2, Pascal VOC 2007; the resulting mAP is 66%. RCNN is
you’ll find a description of the literature review. In Sect. 3, being used as a basis.

123
 Wireless Networks

He et al. [14] propose SPP-Net (Spatial Pyramid Pooling addition to deep convolutional neural networks, more tra-
in Deep Convolutional Networks for Visual Recognition) ditional feature-design approaches were used during the
to improve detection performance. RESNET’s purpose is study. This was done to ensure the accuracy of the material.
to recover the system’s depth to collect features with better (CNN).
expression ability and higher accuracy by addressing the Kvasic et al. [18] employed deep convolutional neural
issue of network migration through the inclusion of a networks (CNNs) to recognize undersea items using sonar
residual module. Support Vector Machine (SVM) is footage. Their investigation revealed that their approach
replaced by Multilayer Perceptron (MLP) in Fast RCNN, produced considerably superior results when compared to
which makes training and classification much better. the current gold standard in technology. Considerations for
Duggal and colleagues [15] The proposed model was to supervised learning techniques, balancing of imbalanced
automatically characterise the movie using object detection data, difficulty of a procedure, the source universe’s com-
methods. Video content explanation is a simple task for plexity, the user’s commotion, differentiated data,
humans but a challenging one for computers. The foun- rebudging information interconnections and exhibit non-
dation of the proposed system is the YOLO object detec- linearities. Researchers developed a deep learning system
tion method. Because it is faster and uses less memory than that can recognize and classify fish species from images, as
the other two models, their proposed model outperforms well as locate the subjects of the photographs.
them. They used the YOLO method to count the fish, A benchmark that was developed by Mossbauer and
which we will change to fit our own needs. colleagues [19] and based on the Singapore Maritime
Ren et al. [16] used Fast RCNN, which includes RPN to Dataset was given (SMD). This collection includes both
choose and change region proposals rather than selective onshore and onboard marine items, as noted in Table 3.
search, to tackle the end-to-end detection problem. At Additionally, it includes visual-optical and near-infrared
ECCV2016 (European Conference on Computer Vision), films as well as comments for item discovery. In this study,
Liu Wei planned a new technique named SSD (Single Shot the authors assess the effectiveness of using two advanced
MultiBox). Utilizing the variable importance attribute of object identification models, namely Faster R-CNN and
the algorithm, users can determine the featured significance Mask R-CNN, in the maritime environment. Recent studies
of each factor in the collection. Every characteristic of the have found evidence to support this assertion [20].
information is given a rating by predictive value; the
greater the value, the more significant or meaningful the
characteristic is to the outcome variable. Its speed advan- 3 Proposed system
tage over faster RCNN stems from its ability to forecast the
organizes and categories of the bounding box without 3.1 Pre-processing
resorting to proposal dispensation.
CNN and other previously developed models of auditory The most significant approach for object detection in
perception were utilized by Yang et al. [9] in order to underwater computer vision is image preprocessing.
detect ship target emitted noise via mel-scale incidence Because of the way light scatters and is absorbed by water,
cepstral constants. The findings demonstrate that deep the images that are produced by the underwater vision
learning may be successfully used to underwater acoustics. system have inconsistent lighting, low contrast, and a sig-
Using the technique of deep learning, the researchers nificant amount of noise [8, 21]. Based on an analysis of
constructed a model for the extraction of underwater current image processing technologies, this study suggests
acoustic sound features. For inflection gratitude, the model ways to improve underwater photos. An advanced form of
was based on a generative hostility system and a deep artificial neural system known as a fully convolutional
neural network classifier. It was designed to extract substitutes the arithmetic operations known as diffusion for
underwater acoustic sound characteristics. Whenever generic mathematical operations in at most one of its
sound moves through one material to the next, its velocity levels.
can alter. Nevertheless, since it behaves resembles a con-
trolled oscillator and preserves the frequencies of the pri- 3.2 The underwater vision detection architecture
mary material, the amplitude typically stays constant. Deep
learning can be implemented in a variety of methods within Illumination, a camera or device, an image acquisition
the realm of submerged auditory goals. card, and application software are common mechanisms of
Li et al. [17] conducted a review on the identification a submerged vision scheme. Images are concentrated
and categorization of optical remote sensing images gen- below the retinal while submerged in water, as opposed to
erated aboard ships. The review concentrated on the on the retina, which causes hypermetropia, which causes an
problem of photograph identification and categorization. In excessively fuzzy vision. Figure 1 depicts the software

123
Wireless Networks

process of an underwater pictorial identification technique. Due to the fact that red and orange light are totally
This process includes image capture, image preprocessing, engrossed at a depth of 10 m in water, images taken
a convolution neural network, and goal detection. underwater typically have a blue-green tint to them. To
During image enhancement and image filtering, it is remove color deviations in underwater photos, color cor-
essential to keep valuable features intact in order to achieve rection must be performed.
the primary goal of picture preprocessing, which is to recover The standard image has very sophisticated color cor-
the contrast of the image while simultaneously minimizing rections applied to it. A broad number of different white-
or eliminating the impact of multiple types of noise to the balancing approaches can be used to correct an image’s
greatest extent possible [22]. Preprocessing is crucial since color deviance so that it more accurately reflects the tem-
pollution will make image processing algorithms inaccurate. perature of the scene. Regardless of the light source or the
The initial step in preprocessing is to compress the picture to entity’s real color, it compels the thing to appear white.
remove noise. The effectiveness of the whole image analysis The Gray Word approach, the max-RGB method, the
cycle depends on how accurately image enhancement is done Shades of Gray method, and the Gray Edge method are just
utilizing filtering. The process of partitioning an image into a few examples of the various methods that fall under this
numerous non-overlapping regions is accomplished with the category. The benefit of Gray Code over binary is that
assistance of a Convolutional Neutral Network. Feature every phase only requires a single bit to modify. Devices
extraction, which seeks to extract the greatest actual key that are susceptible to hiccups as well as other faults will
topographies that best represent the board, is the foundation find this useful [23]. In general, the application scenarios
of thing recognition and organization. Simply said, feature for these technologies involve circumstances that involve
extraction is the process of turning textual information into generic partial color, and the conditions for treatments for
numbers and statistics. To fully visualize the environment in severe underwater vision are not met [24, 25]. The datasets
Text Classification, one relatively simple way to use is image known as components are employed to classify the infor-
enhancement. Because one component is dependent on the mation and preserve it as degrees. Either texts or numbers
others, it is imperative that every effort be made to be suc- can be stored in them. They are helpful in sections with a
cessful. The preprocessing of images and the identification of small variety of distinct entries. Since there is less envi-
targets using underwater vision are the primary focuses of ronmental lighting underneath due to the water’s rapid
this work. Even though dimensional transitions of pictures absorption of sunlight with range, items are harder to see.
(such as spinning, scalability, and transcription) are cate- Additionally, they become blurry due to light refraction
gorized as pre-processing processes because comparable among the item as well as the observer, which further
approaches are employed, the goal of pre-processing is an reduces brightness. This study combines the max-RGB
enhancement of the information that inhibits unintentional method that was developed originally with the shades of
distortions or improves a few attributes crucial for more grey method in order to determine the color that is lighted.
computation. Z
IðxÞ ¼ eðkÞsðk; xÞcðkÞdk ð1Þ
3.3 Combination of max-RGB method
and shades of gray method where I(x) denote the contribution of the underwater
image, e(k) denotes the radiance given off by the light
As a result of the water’s ability to both reflect and absorb source, denotes the wavelength, s(k, x) denotes the sur-
light, colors in images taken underwater tend to be muted. face reflectivity, c(k) signifies the compassion of the

Fig. 1 The processing and object detection structure

123
 Wireless Networks

devices, and w denotes the noticeable range. For the perceived as acoustic energy [7]. Because of the different
examination of submerged infrastructures as well as the shapes of the targets, the echo signals will be different.
identification of numerous man-made items, subsurface Using the information that is contained within the target
picture augmentation is crucial. Improved knowledge of echo signal, we are able to follow the target. An echo is a
marine biology or environmental review are also necessary. mirror of noise that comes to the audience with a latency
The illuminate is defined as following the sound energy in processing of audio signals
Z as well as acoustical. The latency is exactly proportionate
e¼ eðkÞcðkÞdk ð2Þ to the separation between the origin as well as the receiver
x
of the moving mirror. The process of active sonar tracking
The average reflectivity of the division is gray rendering of a moving target involves making predictions on the
to the Grey-World supposition target’s location and actively searching for it. Active sonar
R
sðk; xÞdx tracking of a moving target entail forecasting the target’s
k¼ R ð3Þ
dx location, searching for it, judging it, and associating the
data. Figure 2 depicts the construction drawing of the
Assuming that k is a fixed rate, the interpretation of DCNN-KF technique suggested in this work. Utilizing a
Eq. (1) in terms of its physical significance can be summed variety of construction pieces, including convolution
up as follows: the experiential picture I(x) can be decom- operation, average pooling, or feature maps, CNN is
posed into the produce of the reflectivity of picture S(x) intended to dynamically as well as able to adapt acquire
and the illumination map e(k). Therefore, improving ima- provides advanced of information via training algorithm.
ges with low-light conditions requires removing low-light As active sonar broadcast signs, we make use of periodic
conditions from the source image; Eq. (3) should be pulse signals, and we deploy a uniform linear array to
relieved in Eq. (1) receive echo signals. By offering a higher signal-to-noise
R Z Z
sðk; xÞdx 1 proportion, antenna array increases spectrum utilization
R ¼ R eðkÞsðk; xÞcðkÞdkdx ð4Þ (SNR). Directional antennas stimulate cell coverage and
dx dx x
capacities by collaborating with other antenna techniques
The lighting by elaborating on the fact that the typical like as communication systems or MIMO. Beam-forming
colour of the whole picture multiplied by a power n and matching filtering are utilized on the incoming echo
R n 1n signal in order to locate and identify the target in the
I dx
ke ¼ R ð5Þ preprocessing stage. Beamforming is a method for
dx
enhancing the signal-to-noise ratio of frequency response,
The aforementioned equation can be adjusted in accor- removing unwanted causes of disturbance, and concen-
dance with the max-RGB technique as follows: trating broadcast data to certain areas. Radar systems fre-
R n  quently employ assessment provides, that send out an
I dx
ke ¼ maxIðxÞ  R ð6Þ information in order while looking for recurring compo-
dx
nents in the reflected wave. After applying an azimuth
The defaulting rate of n is 6, which is defined by the allowance to the sign based on the azimuth of the target
shades of grey technique that was planned by Finlayson. that was recognized, a time–frequency spectrogram of the
Where n can be at all amount between 1 and !. target can be generated. After producing the dataset and
labelling the numerous goals, we feed it into deep convo-
3.4 Kalman filter active tracking based on deep lutional neural networks, which are then utilized to train
convolutional neural network the dataset.
During the second stage of tracking, the resonance signs
Following the discharge of sound waves by the active are biased by various azimuths to produce a time–fre-
sonar, pollutants create varying degrees of reverberation quency range picture of the sign that needs to be noticed.
and sound wave dispersion, which are found in the echo After the prototypical has been qualified, it is used for
signals that are continually received throughout the dura- documentation purposes in order to estimate the detach-
tion. A sonic boom is a disturbance that traverses a med- ment and azimuth of the board. The remainder of this
ium, like steel, liquid, or air. According to definitions, a section goes over each stage in detail.
transmitted signal is ‘‘invisible noise with particular The suggested DCNN-KF method is depicted as a block
intensity for humanity’’ and typically has a frequency diagram in Fig. 2, which may be found below. The pro-
greater than 20 kHz. Modern terminology refers to send cedure for the acquisition of data begins with the collection
and receive ultrasonic pulses that are not meant to be of data and the subsequent extraction of a data frame. Data
pre-treatment is a crucial stage in data analysis since it

123
Wireless Networks

Fig. 2 Proposed diagram of DCNN-KF method

helps to diagnose, anomalies, distortion, and lacking vari- acquired, selected, and then organized. Human beings, for
ables. Such information mistakes will persist without example, have sensory receptors for touch, taste, smell,
information pre-treatment in machine learning, lowering sight, and hearing.
the caliber of data analysis. Following this, the data are sent The data is then optimized using the DCNN-KF algo-
to the preprocessing module, where they are subjected to rithm. Deep convolutional neural networks, also known as
additional processing. It is a separate acceleration that is DCNNs, are the type of network that is utilized the most
not a part of other Ascend AI processing components. It is frequently to recognize patterns in digital images and
in charge of carrying out the encrypting, decryption, and motion pictures. DCNNs are the next step in the evolution
preparation operations for pictures or videos. Any type of of old-style artificial neural systems. They use a three-D
processing that is carried out on data that has not yet been neural design that is modelled after the visual brain of
processed in order to get it ready for subsequent data animals. The Kalman filtering process provides estimates
processing is referred to as data preprocessing. This aspect of unseen variables by using measurements that have been
of data preparation is a subset of data preparation. It is acquired over a period of time as its basis. Due to its
often considered to be the most important first step in the extreme precision, CNNs are utilized for picture catego-
process of data mining. The perception module includes rization and identification. It was influenced by how
object detection with Max-RGB and the shade of grey humans recognize objects visually. In conclusion, the data
method, where perception is the process by which sensory are reviewed so that its usefulness may be demonstrated.
information captured in the real world is interpreted,

123
 Wireless Networks

3.4.1 Deep convolutional neural network model with a minimum resolution of 71 by 71 pixels are
supported by Xception’s minimal input layer.
Deep learning buildings are typically categorized as Deep • To achieve great performance at minimal computa-
Belief Networks (DBN), Boltzmann Machines (BM), tional cost compared to other models, Inception-
Restricted Boltzmann Machines (RBM), Deep Auto- ResNet-V2 combines the ideas of the Beginning
Encoders (DAE), and Convolutional Neural Networks. prototypical and the ResNet prototypical. In order to
(DBN stands for Deep Belief Networks; BM stands for get good performance, the Inception model usually
Boltzmann Machines; RBM stands for Restricted Boltz- creates deeper layers. Resnet models excel at training
mann Machines) (CNN). The deep convolutional neural very deep architectures because they apply the concept
network (CNN) has conventional a lot of consideration in of the residual block to the inherent importance of data.
recent research because of its superior ability in learning Data can be acquired in four different ways: through
features from images, which is used for image classifica- gathering updated information, translating or modifying
tion applications [26]. For the purpose of this investigation, historical data, exchanging or trading data, and buying
a number of different deep CNN models were chosen to data. This covers automatic data gathering (for instance,
use. These choices were made with consideration given to through sensors), manually collecting observational
the learning parameters, model layer, computational cost, data, and receiving current information from external
and presentation of deep CNN representations. As a direct resources. The Inception model retains the computa-
result of this, MobileNet [27], MobileNet V2 [28], Incep- tional efficiency it had before combining with the
tion V3 [29], Xception [30], and Inception-ResNet-V2 Resnet model and gains all the benefits of the latter. The
were picked as the winning candidates. Inception-ResNet-V2 model requires a minimum input
layer size of 75 9 75 pixels.
• MobileNet—MobileNet is meant for usage in low-cost
applications and is built for a very compact, low-latency CNN is a subset of the broader family of feed forward
model. The simplified architecture on which Mobile- ANNs; in contrast to fully connected ANNs, in which all
Nets is built employs depth-wise separable convolu- neurons in a given layer are associated to entirely nerve
tions. The goal is to create a deep neural network with cell in the layer below it, its constituent neurons are only
fewer parameters and less impact on the computing interconnected with one another locally. The visual cortex
budget [31]. Among the many uses for MobileNet are of the human brain served as inspiration for this design. For
glocalization on a massive scale, face characteristics, a CNN, the major processing step is a convolution of the
object detection, and face embedding. MobileNet’s filters with a given input image, where the filters them-
lowest input layer can handle 32 9 32 pixel images. selves are represented by a series of arbitrary initialised
• MobileNetV2—Inverted residual with a linear bottle- connections that form the local connection. The structure is
neck is the novel approach that allowed for the made up of many features extraction levels. Convolution,
development of the new mobile architecture known as non-linear neuron activation, and feature pooling are the
MobileNetV2. Because of this, its presentation is three primary building blocks of each stage. The central
increased even further, and it has a lower requirement component of a CNN is the convolutional neural network,
for the primary recollection that the hardware supplies. which is also wherein the preponderance of processing
MobileNetV2 takes 32 9 32 pixel images at the bare takes place. It needs model parameters, a filtering, or a
minimum input layer. Inception V3—The Inception softmax layer, among other things. The activation function
architecture was established by Inception V1, also is significant because it enables a specific system to adapt
known as Google Net. Improvements were made to and carry out complex assignments. In addition, a deep
Google Net’s InceptionV3 by including factorization in neural system can be built with numerous levels of neurons
the model’s architecture. Due to a smaller parameter stacked on top of each other, that is essential for under-
set, Inception requires less processing power than standing large amounts of data with extreme precision. In
VGGNet. Images with a minimum resolution of order to attain satisfactory accuracy in picture categoriza-
75 9 75 pixels can be fed into InceptionV3’s minimum tion, information sharing is a crucial method. Nevertheless,
input layer. the majority of aggregating techniques are intuitive as well
• Xception, short for ‘‘Extreme Inception,’’ is a more as uncontrolled [32]. A convolutional neural network
effective alternative to Inception V3. Xception’s archi- (CNN) is depicted in its most fundamental form in Fig. 3.
tecture is inspired by Inception’s, with the latter’s When many successive steps of feature extraction are
Inception modules being swapped out for depth wise linked together in a CNN, we say that the CNN is deep.
separable convolutions to achieve better model effi-
ciency and better results than Inception V3. Images

123
Wireless Networks

3.5 Kalman filter In addition to this, it is essential to model either the pro-
cess of measuring or the connection that exists among the
The Kalman filter has been the subject of numerous liter- state and the quantity. It is not always feasible to immediately
ature-based derivations. One might use a variety of terms see the process, which is more of a general statement (i.e.,
to describe the change from one condition to another. There is no room for error in the observability of any of the
Generally speaking, we may classify these possibilities as state parameters). It’s possible that some of the parameters
either linear or non-linear functions describing the state representing the state can’t even be observed in any way; the
transition. Both types of transitions are manageable, parameters that are measured may be scaled parameters or
although the traditional Kalman filter uses a linear transi- they may be a composite of several parameters. Once more,
tion function. Non-linearity in both the measurement we are making the supposition that the association is linear.
relationship and the transition itself is possible with the Therefore, the value of the dimension zk can be written as an
help of the extended Kalman filter (EKF). The traditional expression of the state x k using.
Kalman filter can use the equation that shows how state z k ¼ H xk þ v k ð9Þ
k - 1 changes into state k Fig. 4.
where H is a m 9 n matrix that describes the relationship
xk ¼ Axk1 þ wk1 ð7Þ
between the measurement and the state. In a manner very
where A is the matrix that reflects the state changes and similar to that of the process, vk - 1 is the noise associated
w_(k - 1) is a term that represents sound. Because this with the measurement. In addition to this, it is presumed to
noise period is a Gaussian random variable with a zero have a normal distribution, which may be represented as
mean and a covariance matrix Q, the probability circulation pðvÞ  Nð0; RÞ ð10Þ
of its standards appearances like this:
where R is the covariance matrix referred to as measure-
pðwÞ  Nð0; QÞ ð8Þ
ment noise covariance matrix.
For the remainder of this investigation, we will refer to In our examination, the state xk covers the location (x, y)
the covariance matrix that is indicated by the letter Q as the of the object at the instant k and also the speed of the thing
process noise covariance matrix. It takes into consideration in both x_ and y_ instructions. The new location (xk ,yk ) is the
the possibility that the process could alter between step old location (xk1 ,yk1 ) plus the speed ðx_k1 ; y_k1 Þ plus
k - 1 and step k in ways that are not previously taken into noise w k - 1.
account in the state transition matrix. Another quality that The equation for the state, Eq. (11), is defined as follows
is presumed to be possessed by w_(k - 1) is its indepen- when applied to this scenario:
dence from the state x_ (k - 1).

Fig. 3 Architecture of deep convolution neural network

123
 Wireless Networks

Fig. 4 Estimation and monitoring of target characteristics using the Kalman filter

2 3 2 2 3 forecasting enables firms to create fair and quantifiable

xk 3 xk1
10t0 6 y 7 goals. The following is an outline of the most important
6 yk 7 6 6 k1 7
6  7 6 010t 7 76 x 7 þ wk1 parts of the Kalman Filtering method that was used for this
6 xk 7 ¼ 4 ð11Þ
4  5 0010 56
4 k1
7
y  5 paper:
0001 y
k k1
Time Update (Predict) equations:
Step 1: Project the state ahead:
where ‘t’ stands for the amount of time that has passed
xbk  ¼ Ab
x k1 þ wk ð13Þ
between any two consecutive image frames and is regarded
to be one second. And the extent equation Eq. (12) is Step 2: The error covariance will be projected in
written as advance:
2 3
xk pk  ¼ APk1 AT þ Q ð14Þ
    6 yk 7
x 1000 6 7
6 x 7 þ vk Measurement update (Correct) equations:
zk ¼ kmeas ¼ ð12Þ
ykmeas 6
0100 4 k 7 Step 3: Compute the Kalman gain:
 5
y
k K k ¼ P H T ðHP K H T þ RÞ1 ð15Þ

where xkmeas and ykmeas are the restrained positions in x and Step 4: Adjust the estimates based on the measurements:
y instructions. xb ¼ xbk  þ K k ðzk  H xbk  Þ ð16Þ
By making an estimate of the state of the procedure at
approximately point in period and then receiving response Step 5: Keep the error covariance up to date:
from (noisy) measurements, the Kalman filter can provide Pk ¼ ðI  K k HÞPk  ð17Þ
an approximation of the process’s behavior. So, the Kal-
Step 6: Go to Step 1.
man filter calculations can be broken down into two cate-
The present state and error covariance (pk  ) approxi-
gories: those that update over time and those that update
mations have to be projected forward (in time) in order to
across space and time. For the next time step, the a priori
get the a priori (bx k  ) estimations for the subsequent time
estimates of the state and fault covariance are pre-
step, and steps 1 and 2 are accountable for this projection.
dictable frontward (in time) by the period update calcula-
Feedback, or the inclusion of a new dimension into the a
tions. The feedback, or the incorporation of a new extent
priori approximation in order to achieve a better a poste-
into the priori approximation in order to produce a more
riori approximation, is the accountability of Steps Three
accurate posteriori approximation, is the responsibility of
through Five in the process (b x k  ). In (step 3), the Kalman
the measurement update computations.
gain K k is determined to be the gain that results in the least
The period informs projects the present state approxi-
amount of variation in the posteriori error (Eq. 15). The
mation into the future, as shown in Fig. 5, but the extent
following stage is to amount the procedure in order to get
update adjusts the predictable approximation based on a
zk , and then an a posteriori estimate of the state is generated
current quantity. The calculations that are used to inform
by inserting the quantity into the calculation in the manner
the period can also be thought of as being forecaster cal-
described above (16). The very last thing that has to be
culations, while the calculations that are used to update the
done is to calculate a posteriori error covariance through
capacities can be thought of as corrector calculations. The
the Eq. (17).
advantages of using forecaster calculations are assists in
The process is repeated after each set of updated times
predicting market shift, aids in planning and goal setting, as
and measurements, with the past posteriori estimates being
well as aids in budgeting. Using past and present data,

123
Wireless Networks

Fig. 5 Kalman filter cycle

used to anticipate the current set of prior estimates. The • True Positives (TP): The incidence that we predicted to
fact that the Kalman filter is recursive is a major selling take place did, in fact, take place, and the actual yield
point for it. In comparison to (say) a Wiener filter imple- was also accurate.
mentation, which is meant to work on entirely the infor- • True Negatives (TN): The incidence for which we
mation straight for apiece approximation, this makes actual expected a false yield, and the actual yield turned out to
implementations considerably more viable. The Kalman be false as well.
filter, on the other hand, uses all of the data from the past to • False Positives (FP): The incidence that we predicted to
shape the estimate for the present. be true did not take place, and the actual produce was
also inaccurate.
• False Negatives (FN): The instance in which we
4 Result and discussion predicted the yield to be fake but it actually turned
out to be true as well.
4.1 Evaluation metrics • Precision: To determine it, divide the entire amount of
correct positive outcomes by the entire amount of
To accomplish the excellent heftiness of the suggested optimistic results that the classifier expected.
method, a number of DL models are used. In this work, we TP
assess the capabilities of various DL models employing Precision ¼ ð18Þ
TP þ FP
various indoor localization methodologies. The four
parameters of true positives (TP), true negatives (TN), false Recall: It is the ratio of the entire amount of positive
negatives (FN), and false positives (FP) are crucial to the outcomes that are accurate to the total amount of examples
evaluation of the measure (FP). Measures including accu- that are conjugate (entirely examples that ought to have
racy, recall, precision, F-score, root-mean-squared error, been recognized as sure).
and runtime are used to evaluate a classification system’s
TP
efficiency. The Accuracy of Classification Criteria (ACC) Recall ¼ ð19Þ
is used to amount the proportion of right classifications. TP þ FN
Exactness measures how closely a predicted value matches F1-score: It is also known as the harmonic mean, and its
the genuine value, and Se measures the proportion of true goal is to strike a compromise between accurateness and
positives. Harmonically averaging precision and recall recall. It is able to calculate with both false positives and
yields the F1-score. As a result, it is a more general false negatives, and it purposes efficiently on a dataset that
approach to achieving equilibrium in Pr. The root-mean- is excessive.
squared error (RMSE) is an mistake statistic that provides a 2TP
cumulative approximation of mistake. The statistic is F  Score ¼ ð20Þ
2TP þ FP þ FN
intended by taking the four-sided root of the sum of the
squared mistakes present in our data set. It gives a single Accuracy: It refers to the proportion of accurate fore-
metric to evaluate efficiency regardless of the classification casts relative to the total amount of examples that were fed
cutoffs used. The following table defines each of these into the model.
measures.

123
 Wireless Networks

TP þ TN DCNN-KF technique obtains an F-Score of 90.13%,

Accuracy ¼ ð21Þ
TP þ TN þ FP þ FN whereas the RCNN method achieves 78.17%, the CNN
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
iffi method achieves 81.23%, the Faster RCNN method
1 Xk h 2 2
RMSE ¼ ðbx j  xbj Þ þ ðb y j  ybj Þ ð22Þ achieves 84.39%, and the YOLOv3 method reaches
K j¼1
87.31% of the F-score. Using 500 data points, the proposed
method obtains a 92.63% F-Score when compared to
4.2 Mean average precision analysis 80.83%, 83.81%, 86.28%, and 89.51% of F-score achieved
by RCNN,CNN, Faster RCNN, and YOLOIv3 respec-
Table.1 and Fig. 6 provide a comparison of the DCNN-KF tively. It is reasonable to conclude that the DCNN-KF
method’s precision to that of a different method already in method is superior.
use. The findings prove that the proposed strategy outper-
forms the current best practices. For 100 data points, the 4.5 Accuracy analysis
DCNN-KF technique obtains a precision of 92.17%,
whereas the RCNN method achieves 82.33%, the CNN Table 4 and Fig. 9 provide a comparison of the DCNN-KF
method achieves 85.31%, the Faster RCNN method method’s accuracy to that of other methods already in use.
achieves 87.25%, and the YOLOv3 method reaches The findings prove that the proposed strategy outperforms
89.46% of precision. Using 500 data points, the proposed the current best practises. For 100 data points, the DCNN-
method achieves a precision of 94.55%, compared to KF technique obtains an accuracy of 94.25%, whereas the
84.78%, 87.03%, 89.12%, and 91.28% for RCNN, CNN, RCNN method achieves 82.13%, the CNN method
Faster RCNN, and YOLOv3, respectively. It is reasonable achieves 85.31%, the Faster RCNN method achieves
to conclude that the DCNN-KF method is superior. 87.18%, and the YOLOv3 method reaches 90.61% of
accuracy. Using 500 data points, the proposed method
4.3 Recall analysis obtains a 96.36% accuracy when compared to the 84.19%,
86.78%, 89.85%, and 91.69% of accuracy achieved by
Table 2 and Fig. 7 provide a comparison of the DCNN-KF RCNN, CNN, Faster RCNN, and YOLOIv3 respectively. It
method’s recall to that of different methods already in use. is reasonable to conclude that the DCNN-KF method is
The findings prove that the proposed strategy outperforms superior with more accuracy.
the current best practises. For 100 data points, the DCNN-
KF technique obtains a recall of 85.63%, whereas the 4.6 RMSE analysis
RCNN method achieves 74.33%, the CNN method
achieves 77.28%, the Faster RCNN method achieves The RMSE comparison of the DCNN-KF approach with
80.15%, and the YOLOv3 method reaches 82.42% of current methods is shown in Table 5 and Fig. 10. The
precision. Using 500 data points, the proposed method outcomes demonstrate that the suggested technique is
achieves an 87.19% precision while RCNN, CNN, Faster greater to the other approaches. Using 100 data points, the
RCNN, and YOLOv3 achieve 76.78%, 79.47%, 81.78%, RMSE for the DCNN-KF approach is 64.31%, whereas it is
and 84.57% recalls, respectively. It is reasonable to con- 75.35% for the RCNN method, 72.13% for the CNN,
clude that the DCNN-KF method is superior. 70.16% for the Faster RCNN method, and 66.94% for the
YOLOv3. Using 500 data points, the RMSE for the sug-
4.4 F-score gested technique is 66.46%, whereas it is 77.93%, 74.64%,
71.53%, and 68.79% for RCNN, CNN, Faster RCNN, and
Table 3 and Fig. 8 provide a comparison of the DCNN-KF YOLOv3 respectively. This demonstrates the greater pre-
method’s F-Score to that of different methods already in sentation of the planned method with a lower error rate.
use. The findings prove that the proposed strategy outper-
forms the current best practises. For 100 data points, the

Table 1 Mean average

No of data from dataset RCNN CNN Faster RCNN YOLOv3 DCNN-KF
precision analysis of DCNN-KF
method with existing system 100 82.33 85.31 87.25 89.46 92.17
200 83.46 86.48 87.69 90.16 92.69
300 83.81 85.81 88.16 90.73 93.41
400 84.24 86.83 88.83 89.82 93.75
500 84.78 87.03 89.12 91.28 94.55

123
Wireless Networks

Fig. 6 Mean average precision analysis of DCNN-KF method with existing system

Table 2 Recall analysis for

No of data from dataset RCNN CNN Faster RCNN YOLOv3 DCNN-KF
DCNN-KF method with
existing system 100 74.33 77.28 80.15 82.42 85.63
200 74.77 78.49 80.64 83.47 85.98
300 75.39 77.59 81.23 82.79 86.16
400 76.47 78.81 82.11 83.71 86.84
500 76.78 79.47 81.78 84.57 87.19

Fig. 7 Recall analysis of DCNN-KF method with existing system

123
 Wireless Networks

Table 3 F-Score analysis of

No of data from dataset RCNN CNN Faster RCNN YOLOv3 DCNN-KF
DCNN-KF method with
existing system 100 78.17 81.23 84.39 87.31 90.13
200 78.93 81.81 84.71 88.64 91.47
300 79.61 82.36 85.18 87.82 91.83
400 80.61 83.47 85.79 88.91 90.78
500 80.83 83.81 86.28 89.51 92.63

Fig. 8 F-score analysis for DCNN-KF method with existing system

Table 4 Accuracy analysis of

No of data from dataset RCNN CNN Faster RCNN YOLOv3 DCNN-KF
DCNN-KF method with
existing system 100 82.13 85.31 87.18 90.61 94.25
200 82.69 86.45 87.58 90.92 94.77
300 83.38 86.81 88.51 91.11 95.48
400 83.78 85.94 88.83 92.78 95.73
500 84.19 86.78 89.85 91.69 96.36

5 Conclusion problematic for underwater pictures is planned. This

method is used to train the plotting association in order to
The characteristics of vision while submerged are investi- get the lighting map. Following the completion of picture
gated in this study. Image processing has advanced to the dispensation, a DCNN-KF technique for underwater
point where there are new approaches that can address detection and classification is constructed. Following this,
issues such as poor contrast and inadequate lighting. The two modified algorithms are utilized in order to alter the
deep DCNN-KF strategy, which is widely acknowledged as DCNN-KF structure according to the characteristics of
the method for detecting items that is the quickest overall, underwater vision. When compared to the detection results
is presented as an option for the detection and categoriza- of the other schemes, scheme 2’s results clearly show that
tion of marine species. After improving submerged image it provides greater detection. Even though some things are
with a grouping of the max-RGB technique and the shade misused, the efficiency and capabilities of the suggested
of grey technique, a deep convolutional neural network technique are clearly confirmed by the results of the
(DCNN-KF) technique for resolution the weakly illumined qualitative and measurable evaluations. The trained model

123
Wireless Networks

Fig. 9 Accuracy analysis of DCNN-KF method with existing system

Table 5 RMSE analysis of

No of data from dataset RCNN CNN Faster RCNN YOLOv3 DCNN-KF
DCNN-KF method with
existing system 100 75.35 72.13 70.16 66.94 64.31
200 75.81 72.59 69.15 67.14 65.17
300 76.16 73.19 70.67 67.69 65.84
400 77.52 73.71 69.82 68.42 64.72
500 77.93 74.64 71.53 68.79 66.46

Fig. 10 RMSE analysis for DCNN-KF method with existing system

123
 Wireless Networks

is utilized to aid the object in detecting submerged objects. Communications and Mobile Computing, 2022, 1–12. [Link]
In spite of the fact that the planned technique is appropriate org/10.1155/2022/2086613
7. Lakshmanna, K., Subramani, N., Alotaibi, Y., Alghamdi, S.,
for our submerged automaton to use for item detection, it is Khalafand, O. I., & Nanda, A. K. (2022). Improved metaheuris-
not superior to the usual approaches that were used for the tic-driven energy-aware cluster-based routing scheme for IoT-
other datasets. The method DCNN-KF that was proposed assisted wireless sensor networks. Sustainability, 14, 7712.
in this work was successful in accomplishing what past [Link]
8. Noh, J.M., Jang, G.R., Ha, K.N., Park, J.H. (2019). Data aug-
tracking systems were unable to do: it reduced the number mentation method for object detection in underwater environ-
of calculations that were required by the system, improved ments. In Proceedings of the 19th international conference on
tracking accuracy, and greatly decreased tracking devia- control, automation and systems (pp. 324–328), Jeju, Korea.
tions. In the future, it will be important to investigate how 9. Yang, H., Shen, S., Yao, X., Sheng, M., & Wang, C. (2018).
Competitive deep-belief networks for underwater acoustic target
factors such as velocity, acceleration, and other variables recognition. Sensors, 18, 952.
impact the EKF-based localization method. In the future, 10. Yao, X.H., Yang, H.H., Li, Y.Q. (2019). A method for feature
efforts will be focused on improving the localization extraction of hydroacoustic communication signals based on
algorithm that is based on EKF and introducing two generative adversarial networks. In Proceedings of the 2019
academic conference of the underwater acoustics branch; Nan-
additional localization methods, including the UKF and PF jing, China, Chinese Society of Acoustics: Beijing, China.
localization algorithms. These are the two localization 11. Kumar, D. R. (2021). Hybrid unscented Kalman filter with rare
methods that will be introduced. features for underwater target tracking using passive sonar
measurements. Optik, 226, 165813.
12. Lamyae, F., Siham, B., & Hicham, M. (2021). Mathematical
model and attitude estimation using extended colored Kalman
Authors’ contributions All author is contributed to the design and
filter for transmission lines inspection’s unmanned aerial vehicle.
methodology of this study, the assessment of the outcomes and the
IIETA, 54, 529–537.
writing of the manuscript.
13. Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). ImageNet
classification with deep convolutional neural networks. Commu-
Funding Authors did not receive any funding.
nications of the ACM, 60(6), 84–90.
14. Girshick, R. (2015). Fast R-CNN, In 2015 IEEE International
Data availability No datasets were generated or analyzed during the
Conference on Computer Vision (ICCV), (pp. 1440–1448),
current study.
Santiago, Chile.
15. He, K., Zhang, X., Ren, S., & Sun, J. (2015). Spatial pyramid
pooling in deep convolutional networks for visual recognition.
Declarations IEEE Transactions on Pattern Analysis and Machine Intelli-
gence, 37(9), 1904–1916.
Conflict of interest Authors do not have any conflicts. 16. Duggal, S., Manik, S., Ghai, M. (2017). Amalgamation of video
description and multiple object localization using single deep
learning model, In Proceedings of the 9th international confer-
References ence on signal processing systems (pp. 109–115). New York,
USA: ACM.
17. Li, B., Xie, X., & Wei, X. (2020). Ship detection and classifi-
1. Jalal, A., Salman, A., Mian, A., Shortis, M., & Shafait, F. (2020). cation from optical remote sensing images: a survey. Chinese
Fish detection and species classification in underwater environ- Journal of Aeronautics, 34, 145–163.
ments using deep learning with temporal information. Ecological 18. Kvasic, I., Miškovic, N., Vukic, Z. (2019). Convolutional neural
Informatics, 2020(57), 101088. network architectures for sonar-based diver detection and track-
2. Veeramani, T., Bhatia, S., & Memon, F. H. (2022). Design of ing. In Proceedings of the OCEANS 2019 (pp. 17–20). Marseille,
fuzzy logic-based energy management and traffic predictive France.
model for cyber physical systems. Computers and Electrical 19. Moosbauer, S., Konig, D., Jakel, J. (2019). A benchmark for deep
Engineering, 102, 108135. [Link] learning-based object detection in maritime environments, In
eng.2022.108135 Proceedings of the IEEE/CVF conference on computer vision and
3. Zhu, B., Wang, X., Chu, Z., Yang, Y., & Shi, J. (2019). Active pattern recognition workshops, Long Beach, CA, USA.
learning for recognition of shipwreck target in side-scan sonar 20. Huang, H., Zhou, H., Yang, X., Zhang, L., Qi, L., & Zang, A.-Y.
image. Remote Sensing, 11, 243. (2019). Faster R-CNN for marine organisms’ detection and
4. Yang, H., Byun, S.-H., Lee, K., Choo, Y., & Kim, K. (2020). recognition using data augmentation. Neurocomputing, 337,
Underwater acoustic research trends with machine learning: 372–384.
Active SONAR applications. Journal of Ocean Engineering and 21. Girshick, R., Donahue, J., Darrell, T., Malik, J., (2014). Rich
Technology, 34, 277–284. feature hierarchies for accurate object detection and semantic
5. Nguyen, H.-T., Lee, E.-H., & Lee, S. (2019). Study on the segmentation, In 2014 IEEE conference on computer vision and
classification performance of underwater sonar image classifica- pattern recognition (pp. 580–587), Columbus, OH, USA.
tion based on convolutional neural networks for detecting a 22. Ghani, A. S. A., & Isa, N. A. M. (2015). Enhancement of low-
submerged human body. Sensors, 20, 94. quality underwater image through integrated global and local
6. Sreekala, K., Cyril, C. P. D., Neelakandan, S., Chandrasekaran, contrast correction. Applied Soft Computing, 37, 332–344.
S., Walia, R., & Martinson, E. O. (2022). Capsule network-based 23. Muschelli, J. (2020). ROC and AUC with a binary predictor, a
deep transfer learning model for face recognition. Wireless potentially misleading metric. Journal of Classification, 37(3),
696–708.

123
Wireless Networks

24. Anuradha, D., Khalaf, O. I., Alotaibi, Y., Alghamdi, S., & refereed international journals and 10 research papers in the pro-
Rajagopal, M. (2022). Chaotic search-and-rescue-optimization- ceedings of various international conferences. Her areas of research
based multi-hop data transmission protocol for underwater include Machine Learning, Artificial Intelligence and Grid
wireless sensor networks. Sensors, 22, 2867. [Link] Computing.
3390/s22082867
25. Alotaibi, Y., Alghamdi, S., & Khalaf, O. I. (2022). An efficient Dr. N. Nijil Raj , currently
metaheuristic-based clustering with routing protocol for under- working as professor and head,
water wireless sensor networks. Sensors, 22(2), 415. [Link] department of CSE, Younus
org/10.3390/s22020415 College of engineering and
26. Xu, Y., Zhang, Y., Wang, H., Liu, X. (2017). Underwater image technology, Kollam, affiliated to
classification using deep convolutional neural networks and data APJ technological university,
augmentation. In Proceedings of the 2017 IEEE international Thiruvananthapuram, Kerala,
conference on signal processing, communications and computing He has published more than 15
(ICSPCC), Xiamen, China. papers in national and interna-
27. Howard, A.G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., tional journals. He was post
Weyand, T., Andreetto, M., Adam, H. (2017). MobileNets: effi- graduated from M. S. Univer-
cient convolutional neural networks for mobile vision applica- sity, Tamilnadu in [Link].,
tions. arXiv:1704.04861. MCA. and MBA. from MG
28. Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., Chen, L.C. University Kottayam, Kerala.
(2018). MobileNetV2: Inverted residuals and linear bottlenecks. His area interest is bioinfor-
In Proceedings of the 2018 IEEE/CVF conference on computer matics, Machine Learning and AI, and Image Processing. He is
vision and pattern recognition (pp. 4510–4520), Salt Lake City, having 18 years of teaching experience in UG and PG level. He has
UT, USA. completed the Doctoral Degree from M. S. University, Tamilnadu, in
29. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z. Computer Science and Information Technology.
(2016) Rethinking the inception architecture for computer vision.
In Proceedings of the 2016 IEEE conference on computer vision Dr. Sachi Gupta , Professor and
and pattern recognition (CVPR) (pp. 2818–2826.), Las Vegas, Head of the IT Department at
NV, USA. IMS Engineering College,
30. Chollet, F. Xception: Deep learning with depthwise separable Ghaziabad, has more than
convolutions. In Proceedings of the 30th IEEE conference on 18 years of teaching and
computer vision and pattern recognition (CVPR) (pp. research experience. She com-
1800–1807), Honolulu, HI, USA. pleted her Phd. and [Link].
31. Kaya, A., Keceli, A. S., Catal, C., Yalic, H. Y., Temucin, H., & (gold medalist) degrees from
Tekinerdogan, B. (2019). Analysis of transfer learning for deep Banasthali Vidyapith, Rajasthan
neural network based plant classification models. Computers and in the Computer Science
Electronics in Agriculture, 158, 20–29. domain. She completed her B.
32. Sajjad, M., Khan, S., Muhammad, K., Wu, W., Ullah, A., & Baik, Tech. from UPTU, Lucknow.
S. W. (2019). Multi-grade brain tumor classification using deep She has filed six patents, out of
CNN with extensive data augmentation. Journal of Computer which three have been granted,
Science, 30, 174–182. and published more than twenty
papers in national/international level conferences/ journals of repute.
Publisher’s Note Springer Nature remains neutral with regard to She is an active member of CSI, Vibha, IACSIT, IAENG, etc. She has
jurisdictional claims in published maps and institutional affiliations. also worked as a national and international advisory board member
for various reputed conferences. Her areas of interest include Genetic
Algorithms, Machine Learning, and Fuzzy Logic.
Springer Nature or its licensor (e.g. a society or other partner) holds
exclusive rights to this article under a publishing agreement with the
author(s) or other rightsholder(s); author self-archiving of the Dr. G. Anitha holds a Doctorate
accepted manuscript version of this article is solely governed by the (PhD.) in Information and
terms of such publishing agreement and applicable law. Communication from Anna
University, a Master’s Degree
Dr. Keshetti Sreekala working as
(M.E.-CSE) from Anna
an Assistant Professor in the
University, and Bachelor’s
Department of Computer Sci-
degree (B.E.-CSE) from Madu-
ence and Engineering at
rai Kamaraj University. She is
Mahatma Gandhi Institute of
having 17 years of teaching
Technology (A), has about
experience at Rajalakshmi
18 years of teaching experience.
Engineering College as Assis-
She received her [Link] degree
tant Professor (SG) where she
in Computer Science from
guided many U.G and P.G pro-
Jawaharlal Nehru Technological
jects. Currently, she is working
University, Hyderabad. She
in Amrita Vishwa Vid-
received Ph.D. degree in Com-
hyapeetham as Assistant Professor. She has participated in many
puter Science Engineering from
International conferences and published research articles in peer-re-
Jawaharlal Nehru Technological
viewed journals. She is an active researcher in Artificial Intelligence,
University, Hyderabad, Telan-
Machine Learning, Deep Learning and Computer Vision.
gana state. She has published 16 research papers on emerging areas in

123
 Wireless Networks

Dr. Ashok Kumar Nanda is Mr. Abhay Chaturvedi is work-

working as an Associate Pro- ing as Associate Professor in the
fessor in the Department of CSE Department of Electronics &
at the B. V. Raju Institute of Communication Engineering,
Technology (BVRIT), Narsa- GLA University, Mathura, U.P.
pur, Medak, Telangana, India. He has about twenty years of
BVRIT is recognized by AICTE teaching experience at [Link].
and UGC-Autonomous, NBA & [Link]. level. He received
and NAAC-A accredited. He his B.E. degree in Electronics &
was awarded a Ph. D. (CSE) Communication Engineering
degree from NIT Hamirpur, with Hons and M.E. degree in
Himachal Pradesh in May 2015; Microwave Communication &
an M. Tech. (CSE) from GJU, Radar Engineering with Hons
Hisar, Haryana in 1999; and from Dr. B.R. Ambedkar
graduated engineering in Elec- University, Agra, U.P. He is
trical Engineering from The IE (India), Calcutta in [Link] has been pursuing his Ph.D. from the Department of Electronics Engineering,
teaching for more than 21 years. He has published more than 17 Rajasthan Technical University, Kota, Raj. He is the author/co-author
international journals and 10 international conference papers, with a of more than forty research papers at International/National Journals/
total of 12 patents (including international patents) and Indian copy Book-Chapters/Conferences. Three patents are also published with his
rights. His main research interests include lightweight cryptography, name. He is Life Member of IETE New Delhi, ISTE New Delhi,
Information Security, and IoT Security. He is a Fellow Member and ISCA Kolkata and SSI Thiruvananthapuram. He has been awarded
Chartered Engineer of the IE (India), Senior Member of IEEE, with Eminent Professor award by Bestow Edutrex International LLP.
Member of ACM, Life Member of CSI, ISTE, and CRSI.

123

View publication stats