biology

Article
KRS-Net: A Classification Approach Based on Deep Learning
for Koi with High Similarity
Youliang Zheng 1 , Limiao Deng 2 , Qi Lin 3 , Wenkai Xu 1 , Feng Wang 4 and Juan Li 1, *

1 College of Mechanical and Electrical Engineering, Qingdao Agricultural University, Qingdao 266109, China
2 College of Science and Information Science, Qingdao Agricultural University, Qingdao 266109, China
3 Key Laboratory of Cultivation and High-Value Utilization of Marine Organisms in Fujian Province,
Xiamen 361013, China
4 College of Marine Science and Engineering, Qingdao Agricultural University, Qingdao 266109, China
* Correspondence: [email protected]

Simple Summary: The diversity of fish resources is an important component of biodiversity. As

a branch of fish, the diversity of koi varieties is conducive to improving the genetic quality of
offspring, avoiding inbreeding, and improving their adaptability to the natural environment. The
variety classification of koi is a necessary step to improve the diversity of koi varieties and breeding
quality. The traditional manual classification method of koi variety faces some problems, such as
high subjectivity, low efficiency, and high misclassification rate. Therefore, we studied an intelligent
method of classifying koi variety using an artificial intelligence approach, and designed a deep
learning network model, KRS-Net. The intelligent and nondestructive classification was realized for
13 varieties of koi by using the proposed model, and the accuracy rate was 97.9%, which is higher
than that of the classical mainstream classification network. This study provides a reference for
intelligent classification of marine organisms, and can be extended to the screening and breeding of
other species.
Citation: Zheng, Y.; Deng, L.; Lin, Q.;
Xu, W.; Wang, F.; Li, J. KRS-Net: A Abstract: As the traditional manual classification method has some shortcomings, including high
Classification Approach Based on subjectivity, low efficiency, and high misclassification rate, we studied an approach for classifying
Deep Learning for Koi with High
koi varieties. The main contributions of this study are twofold: (1) a dataset was established for
Similarity. Biology 2022, 11, 1727.
thirteen kinds of koi; (2) a classification problem with high similarity was designed for underwater
https://doi.org/10.3390/biology
animals, and a KRS-Net classification network was constructed based on deep learning, which could
11121727
solve the problem of low accuracy for some varieties that are highly similar. The test experiment of
Academic Editors: Chung-Der Hsiao KRS-Net was carried out on the established dataset, and the results were compared with those of five
and Tzong-Rong Ger mainstream classification networks (AlexNet, VGG16, GoogLeNet, ResNet101, and DenseNet201).
Received: 6 October 2022 The experimental results showed that the classification test accuracy of KRS-Net reached 97.90% for
Accepted: 18 November 2022 koi, which is better than those of the comparison networks. The main advantages of the proposed
Published: 29 November 2022 approach include reduced number of parameters and improved accuracy. This study provides an
effective approach for the intelligent classification of koi, and it has guiding significance for the
Publisher’s Note: MDPI stays neutral
classification of other organisms with high similarity among classes. The proposed approach can be
with regard to jurisdictional claims in
published maps and institutional affil-
applied to some other tasks, such as screening, breeding, and grade sorting.
iations.
Keywords: deep learning; classification; AI (artificial intelligence); object recognition; fish

Copyright: © 2022 by the authors.

Licensee MDPI, Basel, Switzerland. 1. Introduction
This article is an open access article
The diversity of fish resources is an important part of biodiversity and is the basis
distributed under the terms and
of the stable and sustainable development of fisheries [1]. As a branch of fish resources,
conditions of the Creative Commons
koi variety diversity is conducive to improving the genetic quality of offspring, avoiding
Attribution (CC BY) license (https://
inbreeding, and improving the adaptability of koi to the natural environment. To improve
creativecommons.org/licenses/by/
the diversity of ornamental fish, classifying koi varieties is a necessary step in breeding
4.0/).

Biology 2022, 11, 1727. https://doi.org/10.3390/biology11121727 https://www.mdpi.com/journal/biology

Biology 2022, 11, 1727 2 of 16

koi for biodiversity. The culture of ornamental fish has a long history. As a representative
ornamental fish, koi have become increasingly popular since the 1980s. Koi are known
as “living gemstones in water”, “geomantic fish”, and “lucky fish” due to their beautiful
physique, bright colors, and elegant swimming posture [2]. In the traditional biological
breeding process, various genes are selectively gathered to form a new variety through
cross-breeding, which can increase species diversity [3]. In the same way, the varieties of koi
have gradually evolved into Tancho, Hikariutsurimono, Utsurimono, Bekko, Kawarimono,
Taisho Sanshoku, Showa Sanshoku, Asagi, Kohaku, Hikarimoyomono, Koromo, Kinginrin,
and Ogon, for a total of thirteen varieties [4,5]. Koi may produce different varieties of
offspring through cross-breeding, and each female lays about 300,000 eggs, with an average
survival rate of more than 80%. Koi are generally sorted three months after hatching, which
is labor-intensive [6]. Additionally, the prices of different varieties of koi greatly vary, so
classifying koi is an indispensable step in koi breeding, screening, and grade sorting.
With the improvement in breeding technology and the expansion of breeding scales [7],
the market value of koi is increasing. As the price and ornamental value of different koi
varieties greatly vary, the screening and grade sorting of koi have become more and more
important [8]. However, regardless of variety screening or grade sorting, they both involve
classification problems [9,10]. Therefore, many scholars have attached great importance
to the research of classification problems for koi varieties. For example, Peng et al. [11]
performed the division of Kohaku, Taisho, and Showa varieties, and proposed a four-
stage classification method for koi. Song et al. [12] divided three varieties of koi from
the perspectives of color, pattern, and lineage. The classification of koi varieties is still
manually performed at present, which depends on those with skilled experienced with
koi. The traditional manual classification method for koi has disadvantages, such as high
subjectivity, low efficiency, and high work intensity. More importantly, classification errors
often occur due to the different skills of sorting workers and the subjectivity of classification.
Especially when there is high similarity between some varieties of koi, such as Taisho and
Showa, misclassification often occurs. To solve the difficulties caused by traditional manual
classification, it is necessary to research intelligent classification approaches for koi varieties.
With the development of computer technology, artificial intelligence, and improve-
ments in hardware performance, deep learning technology has rapidly developed [13–15].
Recognition technology based on deep learning has especially received attention and
been applied in agriculture [16,17], fisheries [18–20], medical treatment [21–23], and other
fields [24–27]. In the field of fisheries, the existing studies have mainly focused on deter-
mining the freshness of marine animals, intelligently recognizing marine organisms, and
classifying economic marine organisms. For example, Lu et al. [28] proposed a method
of automatically recognizing common tuna and saury based on deep learning, and the
final test accuracy of this method was 96.24%. Rauf et al. [29] proposed a fish species
recognition framework based on a convolutional neural network, and the classification ac-
curacy of cultured fish (e.g., grass carp, carp, and silver carp) was 92%. Knausgrd et al. [30]
proposed a temperate fish detection and classification method based on transfer learning.
The classification network was trained on a public dataset for fish (Fish4Knowledge), and
the species classification accuracy was 99.27%. From the above research, it is possible to
classify koi based on image processing and deep learning technology. The above studies
and their related references all discussed methods of classifying underwater animals with
high similarity among species.
To the best of our knowledge, there is currently no study on intelligent classification
methods for underwater animals with high similarity between varieties. Motivated by
the above discussion, we studied a classification approach for underwater animals with
high similarity among varieties. The main contributions of this study are as follows: (1) A
dataset was created for thirteen kinds of koi. To the best of our knowledge, no koi dataset
has yet been reported. (2) A classification problem with high similarity was proposed for
underwater animals, and a KRS-Net network was constructed to solve this classification
problem for koi with high similarity among varieties. (3) The proposed network could
 similarity among varieties. The main contributions of this study are as follows: (1) A da-
taset was created for thirteen kinds of koi. To the best of our knowledge, no koi dataset
Biology 2022, 11, 1727 has yet been reported. (2) A classification problem with high similarity was proposed3 of for16
underwater animals, and a KRS-Net network was constructed to solve this classification
problem for koi with high similarity among varieties. (3) The proposed network could
extract deeper koi feature information, and could further improve the classification accu-
extract deeper koi feature information, and could further improve the classification accuracy
racy by fusing
by fusing the advantages
the advantages of support
of support vector
vector machine
machine (SVM) (SVM)
and aand a fully
fully connected
connected layer.
layer. The superiority of this proposed approach was verified through comparison
The superiority of this proposed approach was verified through comparison experiments exper-
iments with mainstream
with mainstream networks
networks (AlexNet
(AlexNet [31], VGG16
[31], VGG16 [32], [32], GoogLeNet
GoogLeNet [33],[33], ResNet101
ResNet101 [34],
[34], and DenseNet201
and DenseNet201 [35]). [35]).

2.2.Materials
Materialsand
andMethods
Methods
2.1.Image
2.1. ImageAcquisition
Acquisitionand
andData
DataAugmentation
Augmentation
Inthis
In thisstudy,
study, thirteen kinds
kinds of
ofkoi
koiwere
wereselected
selectedasasthe
theresearch
researchobjects, and
objects, and569 original
569 orig-
images of koi with a resolution of 2400 × 1600 were collected using a digital
inal images of koi with a resolution of 2400 × 1600 were collected using a digital cam- camera (EOS
200D, Canon, Tokyo, Japan). Figure 1 shows representative koi images
era(EOS 200D, Canon, Japan). Figure 1 shows representative koi images of the thirteen of the thirteen
varieties. For
varieties. Forthe
theconvenience
convenienceofofdrawing
drawingand
andcharting,
charting,thethe Hikariutsurimono,
Hikariutsurimono,Taisho
Taisho
Sanshoku, Showa
Sanshoku, Showa Sanshoku,
Sanshoku,and andHikarimoyomono
Hikarimoyomono koi varieties are abbreviated
koi varieties as Hikariu,
are abbreviated as
Taisho, Showa,
Hikariu, Taisho, and
Showa,Hikarim, respectively.
and Hikarim, The dataset
respectively. The in this study
dataset was
in this taken
study wasfrom the
taken
actual
from breeding
the data of koi,
actual breeding and
data of the
koi,method
and theused in this
method usedstudy willstudy
in this also be applied
will also betoap-
the
actual koi breeding and production.
plied to the actual koi breeding and production.

Figure 1. Images of thirteen kinds of koi.

Figure 1. Images of thirteen kinds of koi.
To improve the generalization ability of convolutional neural networks, we used image
To improve the generalization ability of convolutional neural networks, we used im-
augmentation methods including brightness, contrast, chroma, mirroring, rotation, and
age augmentation methods including brightness, contrast, chroma, mirroring, rotation,
horizontal or vertical translation. Generative adversarial network is a kind of efficient aug-
and horizontal
mentation or vertical
method [36,37].translation.
The imagesGenerative
generated adversarial
by the usualnetwork is a kind
augmentation of efficient
methods were
augmentation method [36,37]. The images generated by the usual augmentation
sufficient to meet the training task of our network. Therefore, the generative adversarialmethods
were sufficient
network to meet
was not usedthe trainingthe
to expand taskdataset
of our in
network. Therefore,
this study. the generative
A schematic diagramadver-
of the
sarial network was not used to expand the
effect of data augmentation is shown in Figure 2.dataset in this study. A schematic diagram of
the effect of data augmentation is shown in Figure 2.
Biology 2022,
Biology 11, x1727
2022, 11, FOR PEER REVIEW 4 4ofof16
16

Figure
Figure 2.
2. Effect
Effectof
ofdata
data augmentation.
augmentation.

The
The dataset
dataset had
had1464
1464images
images after
afterimage
imageaugmentation,
augmentation, including
including 1027
1027 images
images in in the
the
training
training set,
set, 294
294 images
images in
in the
the verification
verification set,
set, and
and 143
143 images
images in
in the
the test
test set.
set. Because
Because the the
number of
number of images
images of
of Showa
Showa andand Kohaku
Kohaku was
was sufficient,
sufficient, we
we did
did not
not perform
perform additional
additional
data augmentation
data augmentation processing.
processing. TheThe detailed
detailednumber
numberofofimages
imagesfor
foreach
eachvariety
varietyis is
shown
shown in
Table 1.
in Table 1.

Table 1.
Table Numbers of
1. Numbers of images
images of
of thirteen
thirteen koi
koi varieties.
varieties.

Koi Variety Koi Variety Original Images Augmented

Original Images Images
Augmented Images Images after
Images afterAugmentation
Augmentation
Tancho Tancho 37 37 74
74 111
111
Hikariu Hikariu 42 42 84
84 126
126
Utsurimono Utsurimono 25 25 100
100 125
125
Bekko 22 88 110
Bekko 22 88 110
Kawarimono 21 84 105
Taisho Kawarimono75 21 84
39 105
114
Showa Taisho 104 75 39
0 114
104
Asagi Showa 25 104 100
0 125
104
Kohaku 101 0 101
Asagi 25 100 125
Hikarim 22 88 110
Koromo Kohaku 23 101 0
92 101
115
Kinginrin Hikarim 49 22 88
64 110
113
Ogon Koromo 21 23 84
92 105
115
Total Kinginrin 567 49 897
64 1464
113
Ogon 21 84 105
2.2. KRS-Net
Total Classification
567Approach 897 1464
Based on the AlexNet framework, a KRS-Net classification network was designed to
2.2. KRS-Net
classify Classification
thirteen Approach
kinds of koi. KRS-Net is a classification network that is mainly composed
of a residual
Based onnetwork and SVM.
the AlexNet The residual
framework, network
a KRS-Net is used tonetwork
classification extract the
wasfeatures of the
designed to
object, and SVM realizes the classification of objects. A schematic diagram of the
classify thirteen kinds of koi. KRS-Net is a classification network that is mainly composedproposed
network
of is illustrated
a residual network andin Figure
SVM. 3.The residual network is used to extract the features of the
object, and SVM realizes the classification of objects. A schematic diagram of the proposed
network is illustrated in Figure 3.
Biology 2022,11,
Biology2022, 11,1727
x FOR PEER REVIEW 55 of 16
16

Figure3.3. Schematic
Figure Schematicdiagram
diagramof
ofthe
theproposed
proposedKRS-Net.
KRS-Net.

Based on
Based onthetheAlexNet
AlexNetframework,
framework,the themain
mainstructural
structuralchanges
changeswere
wereas asfollows:
follows:
(1)Replace
(1) Replacethe the original
original local
local response
response normalization
normalization(LRN) (LRN)withwith batch
batch normalization
normalization
(BN).
(BN). Both
Both the
the LRN
LRN layer
layer and
and BN BN layer
layer can
can improve
improve the the network
network generalization
generalization ability
ability
and
and training
training speed,
speed, but
but the
the performance
performance of of the
the latter
latter isisusually
usuallysuperior
superior[37]. [37].Compared
Compared
with
with the
the LRN
LRN layer,
layer,the
theBN BNlayer
layer can
can adapt
adapt to to aalarger
largerlearning
learning rate
rate totofurther
further improve
improve
the
the network
network training
training speed.
speed. At At the
the same
same time,
time, itit improves
improves the the effect
effect of
of the
the regularization
regularization
strategy,
strategy,reduces
reducesthe thedependence
dependenceon onthe
thedropout
dropoutlayer,layer,andand improves
improves thethe anti-disturbance
anti-disturbance
ability.
ability.
(2)
(2) Add
Add eight
eight residual
residual blocks
blocks totothe
thenetwork
networkstructure.
structure. TheThe skip
skip connection
connection in in the
the
residual block can overcome the problem of gradient vanishing
residual block can overcome the problem of gradient vanishing caused by increasing caused by increasing depth
in the network.
depth Therefore,
in the network. multiple
Therefore, residual
multiple blocks blocks
residual were introduced to increase
were introduced the depth
to increase the
of the network and extract deeper koi feature information. In addition,
depth of the network and extract deeper koi feature information. In addition, the difficulty the difficulty of
extracting more subtle koi characteristics is
of extracting more subtle koi characteristics is reduced. reduced.
(3)
(3)Fuse
FuseSVMSVMwithwithfully
fullyconnected
connected (FC)
(FC)layer
layerto to
improve
improve accuracy. Inspired
accuracy. Inspiredby [38], we
by [38],
replaced the softmax layer with SVM to achieve higher generalization
we replaced the softmax layer with SVM to achieve higher generalization model perfor- model performance,
thus
mance,improving the accuracy
thus improving of koi of
the accuracy variety classification.
koi variety To improve
classification. To improve the the
accuracy
accuracy of
classification, the fused SVM with FC was added to the network
of classification, the fused SVM with FC was added to the network framework to realize framework to realize
classification
classification of the softmax
of the softmaxlayer.layer.TheThe fused
fused SVMSVM withwith
FC FC transforms
transforms the nonlinear
the nonlinear clas-
classification problem into a linear classification problem
sification problem into a linear classification problem in high dimensional spacein high dimensional space by
by im-
improving the spatial dimensions of the deep feature information
proving the spatial dimensions of the deep feature information extracted from the FC extracted from the FC
layer.
layer.Therefore,
Therefore,the thecomplex
complexprocessing
processing process
processof of feature
featureinformation
information is is simplified,
simplified, and
and
the classification accuracy for koi is further improved.
the classification accuracy for koi is further improved.
The information flow of KRS-Net is as follows: First, thirteen kinds of koi images
The information flow of KRS-Net is as follows: First, thirteen kinds of koi images
were input into the network after data balance. Second, the koi feature information images
were input into the network after data balance. Second, the koi feature information images
was extracted by convolution and pooling, and the extracted feature information was
was extracted by convolution and pooling, and the extracted feature information was
transmitted to the FC layer. The loss rate was reduced by a gradient descent algorithm,
transmitted to the FC layer. The loss rate was reduced by a gradient descent algorithm,
and the feature vectors of the FC layer were imported into SVM. Finally, the optimal
and the feature vectors of the FC layer were imported into SVM. Finally, the optimal clas-
classification hyperplane was obtained by a kernel function, and the parameters of the
Biology 2022, 11, 1727 6 of 16

network were updated. The process of using the kernel function algorithm to explore the
optimal classification hyperplane is as follows:
We assume that the sample space of the training set is T = {( xi , yi ), i = 1, 2, 3, · · · , l }
where xi ∈ Rn is the input sample set, yi ∈ Rn is the output sample set, and i is the ith
training samples. By adopting the appropriate kernel function K ( xi , x j ) = ϕ( xi )· ϕ( x j )
(i = 1, 2, · · · , n ; j = 1, 2, · · · , n ; i 6= j), sample x can be mapped into a high dimen-
sional space, where ϕ( x ) represents the dimensional transform function of x. The convex
optimization problem with constraints is constructed as:
n
1 n n
W (α)max = ∑α− 2 i∑ ∑ αi α j yi y j K ( xi , x j ) (1)
i =1 =1 j =1

n
s.t. ∑ αi yi = 0 (i = 1, 2, 3, · · · , n), 0 ≤ αi ≤ C (2)
i =1

where W (α)max is the optimized object function; C is the penalty parameter; α is the La-

T
grange multiplier, and its optimal solution is α∗ = (α1∗ , α2∗ , · · · , α∗n , which can be obtained
by Formulas (1) and (2). Furthermore, we construct the hyperplane for classification. The
optimal classification hyperplane is defined as follows:

w∗ · x + b∗ = 0 (3)
n
w∗ = ∑ αi∗ yi xi (4)
i =1
n
b∗ = y j − ∑ αi∗ yi K ( xi , x j ), s.t. 0 ≤ α∗j ≤ C (5)
i =1

where w∗ is the normal vector to the optimal hyperplane, and b∗ is the offset.
Then, the categories of koi varieties can be determined through the classification
decision function, which is defined as:
" #
n
f ( x ) = sign ∑ αi yi K xi , x j + b
∗ ∗


(6)
i =1

3. Experimental Results and Analysis

3.1. Setup of Experiment and Performance Indexes
The training work of KRS-Net proposed in this paper was implemented in MAT-
LAB2020a (MathWorks, Natick, MA, USA). The computer performance parameters for
network training were as follows: CPU: Intel (R) Xeon (R) [email protected] GHz; GPU:
NVIDIA RTX 2080Ti; RAM: 64 G; OS: Windows10 (Lenovo, Hong Kong, China).
To unify the training environment and avoid interference from other factors, the
experiments were conducted under the same conditions.
KRS-Net is trained on the established dataset of koi. The learning rate of network
training was uniformly set to 0.0001 according to [39]. Because the values of batch size and
epoch affect the training effect to a certain extent, we studied the influence of the batch
size and epoch on the effect of the network training under different hyperparameters and
obtained the maximum classification performance of the network. The experimental results
are shown in Table 2.
Biology
Biology2022,
Biology 2022,11,
2022, 11,xx1727
11, FOR
FOR PEER
PEER REVIEW
REVIEW 77 7of
ofof16
16
16

Table
Table2.
Table 2.Experimental
2. Experimentaltest
Experimental testresults
test resultsof
results ofKRS-Net.
of KRS-Net.
KRS-Net.

Hyperparameter
Hyperparameter
Hyperparameter Epoch
Epoch = 25 == 25
Epoch 25 Epoch
Epoch===50
Epoch 50
50 Epoch
Epoch
Epoch== 75
7575
= Epoch
Epoch == 100
Epoch 100
= 100
Batch
Batch Size
Size == 44
Batch Size = 4
93.71%
93.71%
93.71%
96.50%
96.50%
96.50%
97.20%
97.20%
97.20%
96.50%
96.50%
96.50%
Batch
Batch Size
BatchSize = =8= 88
Size 97.90%
97.90%
97.90% 97.20%
97.20%
97.20% 96.24%
96.24%
96.24% 95.80%
95.80%
95.80%
Batch
Batch Size
BatchSize
Size= =16
= 16
16 93.01%
93.01%
93.01% 93.01%
93.01%
93.01% 96.50%
96.50%
96.50% 95.10%
95.10%
95.10%
Batch
Batch Size
Batch Size= 32
Size == 32
32 92.31%
92.31%
92.31% 94.41%
94.41%
94.41% 94.41%
94.41%
94.41% 93.01%
93.01%
93.01%
Batch Size = 64 83.92% 90.91% 93.71% 91.61%
Batch
Batch Size
Size == 64
64 83.92%
83.92% 90.91%
90.91% 93.71%
93.71% 91.61%
91.61%

FromTable
From
From Table2,
Table 2,the
2, thebest
the bestclassification
best classificationtest
classification testaccuracy
test accuracyof
accuracy ofKRS-Net
of KRS-Netwas
KRS-Net was97.90%
was 97.90%when
97.90% whenthe
when the
the
batch
batch size
batch size was
size was set
was set
set toto 8 and
to 88 and the
and the epoch
the epoch was
epoch was set
was set to
set to 25.
to 25. When
25. When
When the the batch
the batch size
batch size was
size was 8 and
was 88 and the
and the
the
epochis
epoch
epoch is25,
is 25,Figure
25, Figure444shows
Figure showsthe
shows thetraining
the trainingand
training andverification
and verificationaccuracy
verification accuracycurves
accuracy curvesof
curves ofKRS-Net.
of KRS-Net.From
KRS-Net. From
From
Figure4,
Figure
Figure we can
4, we can see
seethat
thatthe
that theverification
the verificationcurve
verification curveisis
curve close
is toto
close
close thethe
to training
the curve,
training
training which
curve,
curve, indicates
which
which indi-
indi-
that the
cates
cates thatnetwork
that the had better
the network
network had performance.
had better
better The The
performance.
performance. loss curves
The loss
loss of KRS-Net
curves
curves of in the
of KRS-Net
KRS-Net in training
in the and
the training
training
verification
and
and processes
verification
verification processes
processesare shown
are in Figure
are shown
shown in 5. 5.
in Figure
Figure 5.

Figure4.
Figure
Figure 4. Accuracycurves.
4.Accuracy
Accuracy curves.
curves.

Figure 5. Loss curves.

Figure
Figure 5.
5. Loss
Loss curves.
curves.
To further show the specific koi classification with the proposed approach, 143 images
To
To further
further show
show thethe specific
specific koi
koi classification
classification with
with the
the proposed
proposed approach,
approach, 143
143 im-
im-
were used to test the trained KRS-Net. The real and predicted koi categories are summarized
ages
ages were
were used
used to
to test
test the
the trained
trained KRS-Net.
KRS-Net. The
The real
real and
and predicted
predicted koi
koi categories
categories are
are sum-
sum-
in the form of a matrix in Figure 6. Figure 6 shows that each value is the largest in
marized
marized in in the
the form
form of of aa matrix
matrix in
in Figure
Figure 6.
6. Figure
Figure 66 shows
shows that
that each
each value
value is
is the
the largest
largest
the diagonal line in the same columns, which indicates that the KRS-Net had a better
in
in the
the diagonal
diagonal
classification line
line in
effect. in the
the same
same columns,
columns, which
which indicates
indicates that
that the
the KRS-Net
KRS-Net hadhad aa better
better
classification
classification effect.
effect.
Biology 2022, 11, x FOR PEER REVIEW 8 of 16
Biology 2022, 11, x FOR PEER REVIEW 8 of 16
Biology 2022, 11, 1727 8 of 16

Figure 6. Confusion matrix.

Figure 6. Confusion matrix.
Figure 6. Confusion matrix.
3.2.
3.2.Visualization
VisualizationofofFeatures
Features
3.2. Visualization
We of
We visualized the
visualized Features
the features
features toto make
make thethe extracted
extractedfeatures
featuresmore
moreintuitive
intuitivefrom
fromthree
three
aspects:
aspects: different
We visualized network
different network layers,
the features single
to make
layers, image,
singlethe and
extracted
image, gradient-weighted class
features more intuitive
and gradient-weighted activation
classfrom three
activation
mapping
aspects: (Grad-CAM).
mappingdifferent network layers, single image, and gradient-weighted class activation
(Grad-CAM).
To
mapping reflect the
the features
(Grad-CAM).
To reflect features learned
learned by by the
the convolutional
convolutional neural
neural network
network inin the
the training
training
process,
process, the
thesingle
To reflect the and
single andfusion
features features
learned
fusion featuresbywere extracted
extractedand
andvisualized
the convolutional
were neural for
fordifferent
network
visualized in the network
different training
network
layers
layersof
process, KRS-Net,
ofthe as
single and
KRS-Net, asshown
fusionin
shown inFigure
features 7.
Figurewere
7. extracted and visualized for different network
layers of KRS-Net, as shown in Figure 7.

Featurevisualization
Figure7.7.Feature
Figure visualizationof
ofKRS-Net.
KRS-Net.
Figure 7. Feature visualization of KRS-Net.
In Figure 7, the single feature and fusion feature are visualized for the first convolution
layer (Conv1), first pooling layer (Pool1), and the FC layer of KRS-Net from the shallow to
the deep layer. The fused feature is a new combined feature created on the basis of multiple
single features, and the fused features reflect the correlation information among the single
Biology 2022, 11, x FOR PEER REVIEW 9 of 16

In Figure 7, the single feature and fusion feature are visualized for the first convolu-
Biology 2022, 11, 1727 tion layer (Conv1), first pooling layer (Pool1), and the FC layer of KRS-Net from the shal- 9 of 16
low to the deep layer. The fused feature is a new combined feature created on the basis of
multiple single features, and the fused features reflect the correlation information among
the single Eliminating
features. features. Eliminating the redundant
the redundant information information
caused by caused by the correlation
the correlation between singlebe-
tween single
features features
is helpful for isthehelpful for theclassification
subsequent subsequent decision
classification
making decision
of the making
network.ofFrom the
network.
Figure 7,From
we can Figure 7, we
see that canlow-level
some see that some
imagelow-level image feature
feature information wasinformation
extracted bywas the
convolution
extracted layer
by the and pooling
convolution layer
layer andatpooling
the frontlayerof the network,
at the front ofsuch as koi color;
the network, suchandas
some
koi high-level
color; and some image featureimage
high-level information
featurewas extractedwas
information by the FC layer
extracted byatthe
theFCend of the
layer at
network.
the end of the network.
Different activation
Different activation regions
regions are generated
generated at at each
eachlayer
layerdueduetotothe
thediffering
differingabilities
abilitiesto
extract
to image
extract image features
features at at
each network
each network layer when
layer when thethe
image
imageis input intointo
is input the convolutional
the convolu-
neuralneural
tional network. Furthermore,
network. the features
Furthermore, learned
the features by each
learned layerlayer
by each can be canintuitively seen
be intuitively
by comparing
seen by comparing the activation regions
the activation withwith
regions the original
the original image.
image.To study
To studythe the
intermediate
interme-
processing
diate process
processing of KRS-Net
process for afor
of KRS-Net single image,
a single the activation
image, regions
the activation are successively
regions are succes-
shown for the first convolution layer and eight residual blocks
sively shown for the first convolution layer and eight residual blocks by considering by considering a Showa a
example in Figure 8.
Showa example in Figure 8.

Figure
Figure8.8.Activation
Activationregions
regionsofofKRS-Net
KRS-Netininaasingle
singleimage.
image.

From
FromFigure
Figure 8,8, itit can
can be seen that the the shallow
shallownetwork
networkextracted
extractedthe thesimple
simplefeatures
featuresof
of images,
images, andthe
and theextracted
extractedfeatures
features became
became more
more complex
complex andand abstract
abstract with
with theincrease
the increasein
network
in network depth.
depth.
Grad-CAM[40]
Grad-CAM [40]can canvisualize
visualizeaaregion
regionofofinterest
interestofofananimage,
image,which
whichhelped
helpedtotoun-un-
derstandhow
derstand howthetheconvolutional
convolutionalneural neuralnetwork
networkmakesmakesdecisions
decisionson onthetheclassification
classificationof of
koi. Figure 9 gives the Grad-CAM visualization of the first convolution
koi. Figure 9 gives the Grad-CAM visualization of the first convolution layer and eight layer and eight
residualblocks
residual blocksininKRS-Net
KRS-Netfor forkoi.
koi. In In Figure
Figure 9, the
9, the redred region
region of Grad-CAM
of Grad-CAM provides
provides an
an important
important basisbasis
to maketo make classification
classification decisions
decisions oninput
on the the input
imageimage
for thefornetwork,
the network,
and
andblue
the the blue
regionregion
is the is second
the second part.
part. With With
thethe increase
increase in in network
network depth,the
depth, thered
redregion
region
graduallyfocuses
gradually focuseson onthe thespecial
specialcharacteristics
characteristicsofofthe theobject.
object.Taking
Takingthetheimage
imageofofTancho
Tancho
asan
as anexample
examplein inFigure
Figure9,9,we wecan
cansee seethat
thatthe
thered
redregion
regionof ofthe
theoutput
outputimage
imageof ofresidual
residual
block 4 is relatively scattered, but the red region slowly focuses
block 4 is relatively scattered, but the red region slowly focuses on the round spot on on the round spot onthe
the
head of Tancho with the deepening of the network, which is the most
head of Tancho with the deepening of the network, which is the most obvious feature that obvious feature that
distinguishes this variety from other varieties. As can be seen from Figure 9, the network
could effectively capture the characteristics of each koi variety, so that the classification task
was completed well.
Biology 2022, 11, x FOR PEER REVIEW 10 of 16

distinguishes this variety from other varieties. As can be seen from Figure 9, the network
Biology 2022, 11, 1727 10 of 16
could effectively capture the characteristics of each koi variety, so that the classification
task was completed well.

Figure 9.
Figure 9. Grad-CAM
Grad-CAM visualization
visualization of
of KRS-Net.
KRS-Net.

3.3. Comparative Analysis with Other Classification Networks

To
To verify
verify the
the superiority
superiority of the proposed approach, the test accuracy of the KRS-Net
was
was compared with that
compared with that of
of some
some mainstream
mainstream classification networks such
classification networks such as
as AlexNet,
AlexNet,
VGG16, GoogLeNet, ResNet101, and DenseNet201. To visually display
VGG16, GoogLeNet, ResNet101, and DenseNet201. To visually display the comparison the comparison
results,
results, aa 3D
3Dcolormap
colormapsurface
surfacewas
wasused
usedtoto study
study thethe influence
influence of hyperparameters
of hyperparameters on
on the
the test accuracy of the networks. Figure 10 shows the 3D colormap surfaces
test accuracy of the networks. Figure 10 shows the 3D colormap surfaces of the test accu-of the test
accuracy of KRS-Net
racy of KRS-Net and and
that that ofother
of the the other
five five classification
classification networks.
networks.
As can be seen from Figure 10, the highest test accuracy of AlexNet, VGG16, GoogLeNet,
ResNet101, DenseNet201, and the proposed KRS-Net was 94.41%, 94.41%, 93.71%, 93.71%,
96.50%, and 97.90%, respectively. The results of the comparative analysis showed that the
test accuracy of KRS-Net was 1.4% higher than the highest test accuracy of the other five
classification networks, which proves the superiority of the proposed approach. Notably,
the classification effect of the six networks gradually increased with the decrease in batch
size. For this phenomenon, a specific analysis is provided in Section 4.
Biology 2022, 11, x FOR PEER REVIEW 11 of 16
Biology 2022, 11, 1727 11 of 16

(a) AlexNet (b) VGG16

(c) GoogLeNet (d) ResNet101

(e) DenseNet201 (f) KRS-Net (Ours)

Figure
Figure10.
10.Test
Testaccuracies
accuraciesofofdifferent
differentnetworks.
networks.

As
Thecanaccuracy,
be seen precision,
from Figure 10, and
recall, the highest
F1 were test accuracy
selected of AlexNet,
as performance VGG16, Goog-
evaluation indexes
to further
LeNet, analyzeDenseNet201,
ResNet101, the koi classification
and theperformance of the networks,
proposed KRS-Net whose
was 94.41%, definitions
94.41%, 93.71%,are
as follows
93.71%, [41]:and 97.90%, respectively. The results of the comparative analysis showed
96.50%,
that the test accuracy of KRS-Net was = TP +than
1.4% higher TN the highest test accuracy of the
Accuracy (7)
TP + TN + FP + FN of the proposed approach.
other five classification networks, which proves the superiority
Notably, the classification effect of the six networksTP gradually increased with the decrease
Precision = (8)
in batch size. For this phenomenon, a specific analysisTP + FP is provided in Section 4.
The accuracy, precision, recall, and F1 wereTP selected as performance evaluation in-
Recall =
dexes to further analyze the koi classification performance of the networks, whose defini-(9)
TP + FN
tions are as follows [41]:
2 × Pr × Re
F1 = (10)
Pr TP
+ Re+ TN
Accuracy = (7)
TP + TNtrue
where TP, TN, FP, and FN, respectively, represent + FP + FN true negative, false positive,
positive,
and false negative. Accuracy is an intuitive evaluation index, representing the proportion of
TP
classified=to the total number of koi samples. Precision
the number of koi samples correctlyPrecision (8)
represents the proportion of the number of realTP + FP samples to the number of positive
positive
Biology 2022, 11, 1727 12 of 16

samples predicted by the network. Recall represents the proportion of positive samples
predicted by the network to all positive samples. F1 is a comprehensive evaluation index
based on precision and recall. A larger F1 value indicates better network performance.
The comparative experimental results are shown in Table 3. The proposed KRS-Net
had a better effect when the batch size was eight, as shown in Table 2. So, the batch size
was set to 8 and the epoch was set to 25, 50, 75, or 100 in the experiment. The following
performance evaluation indexes of each network represent the average values of thirteen
koi varieties, but not the performance evaluation index of a single variety.

Table 3. Performance evaluation of networks.

Evaluation
AlexNet VGG16 GoogLeNet ResNet101 DenseNet201 KRS-Net
Index (%)
Epoch = 25
Accuracy 98.17 98.82 98.82 98.71 99.25 99.68
Precision 90.51 92.37 92.72 92.98 95.46 97.90
Recall 87.55 91.96 91.89 91.19 94.83 97.76
F1 86.97 91.58 91.00 90.53 94.78 97.80
Epoch = 50
Accuracy 97.63 99.19 98.92 99.14 99.57 99.57
Precision 90.69 94.97 93.41 95.30 97.80 97.58
Recall 88.25 94.69 92.59 93.99 96.92 97.13
F1 87.10 94.29 91.79 93.47 96.87 97.12
Epoch = 75
Accuracy 98.60 98.87 98.39 98.92 98.92 99.35
Precision 91.98 95.05 94.24 94.54 94.68 96.49
Recall 90.49 93.91 93.43 92.52 92.52 95.84
F1 89.93 92.87 93.01 91.67 91.88 95.90
Epoch = 100
Accuracy 98.81 98.82 98.71 95.91 99.16 99.41
Precision 93.51 95.07 92.24 88.00 96.03 96.40
Recall 91.89 90.71 91.26 72.38 93.92 96.54
F1 91.57 91.36 90.82 74.06 93.23 96.17

It can be seen from Table 3 that when the batch size of KRS-Net was set to 8 and
the epoch was set to 25, the classification accuracy, precision, recall, and F1 were 99.68%,
97.90%, 97.76%, and 97.80%, respectively, which are all higher than those of the other five
classification networks. In addition, we can see that the four evaluation indices of the
network all decreased with the increase in epochs, as shown in Table 3, which may have
occurred because the network gradually generated overfitting with the increase in epoch
number in the subsequent training process.
Remark: there is a kind of fine-grained image classification method (subcategory
image classification method), which we used to divide coarse-grained categories into more
detailed subclasses according to the differences in some special parts among subclasses.
However, the difference between the subclasses of koi lies not only in some special parts,
but also in the shape and position of its body patterns as well as the ratio of red, white,
and black. Considering the above factors, we did not choose a similar algorithm such as a
fine-grained algorithm.

4. Discussion
4.1. Factors Influencing Test Accuracy
Although the test accuracy of the proposed KRS-Net reached 97.90%, there were still
some koi varieties that were misclassified. The reasons for the misclassification may include
the following several aspects: (1) Some images were not clear enough because koi swim
quickly. Additionally, the part of the image containing the typical characteristics may have
been incomplete due to the differences in shooting angle during the data acquisition, which
may have affected the training effect of the network. (2) The water ripples generated by the
Biology 2022, 11, 1727 13 of 16

swimming of koi may have affected the image definition, which resulted in the blurring
and distortion of images, as well as other problems [42]. (3) The cross-breeding of koi
reduces the differences between varieties, resulting in a situation where the offspring are
neither like the characteristics of their mother fish nor the various characteristics of their
father, which poses difficulties for classification.

4.2. Influence of Batch Size on Classification Performance

From the 3D colormap surface in Figure 10, we can see that the batch size in the
convolutional neural network had a greater influence on the classification performance
of the network, which was also obtained in [43]. The test accuracies of six classification
networks all decreased with the increase in batch size. A similar phenomenon was shown
in [44]. This phenomenon may be caused by the larger batch size of training data being not
conducive to parameter updating and optimization. On the contrary, a smaller batch size
may be better for solving the positive and negative cancellation problem of the gradient
update value caused by the differences in sampling.

4.3. Advantages of KRS-Net in Structure

The test accuracy of the proposed KRS-Net was higher than that of AlexNet, VGG16,
GoogLeNet, ResNet101, and DenseNet201 in this study, which was determined by the
advantages of the KRS-Net structure. First, the original LRN is replaced by BN based
on AlexNet architecture, which reduces the complexity of the network and improves the
convergence of the network. Second, the addition of residual block deepens the network,
which can extract deeper information and effectively overcome the gradient vanishing
problem. Third, the fusion of SVM with a FC layer replaces the softmax classifier in AlexNet,
which transforms the original nonlinear classification problem into a linear classification
problem to deal with high-dimensional space so that the test accuracy is further improved.

4.4. Influence of Structure on Training Time and Parameters

To identify the factors affecting the network training time and parameter quantity, we
studied the training time and parameters of the six networks when the batch size was 8
and the epoch is 25, as shown in Table 4. It can be seen from Table 4 that there was a high
similarity in training time between KRS-Net and the lightweight network AlexNet, but the
test accuracy of the former was 4.19% higher than that of the latter. This may be because the
addition of the BN layer improves the training speed and convergence of the network. In
addition, KRS-Net has more network layers and connections but has a smaller network size
and fewer parameters than VGG16 and AlexNet. This may be because the skip connection
structure of the residual network not only overcomes the gradient vanishing problem but
also reduces the network size and parameters.

Table 4. Parameters and training times of six classification networks.

Training Number of Number of Size of Network Parameters

Networks
Time (s) Layers Connections (MB) (M)
AlexNet 1079 25 24 227.00 61.00
VGG16 1112 41 40 515.00 138.00
GoogLeNet 1403 144 170 27.00 7.00
ResNet101 3320 347 379 167.00 44.60
DenseNet201 8864 708 805 77.00 20.00
KRS-Net 1338 71 78 49.70 10.89

4.5. Future Work

The evolution of koi has become more and more complex through many years of
breeding and screening. To date, koi can be divided into thirteen categories in a narrow
sense, but more than 100 subcategories have broadly been bred. If more than 100 varieties
of koi can effectively be classified, the time cost and labor force required for koi breeding
Biology 2022, 11, 1727 14 of 16

will be further reduced to a certain extent. Therefore, a multi-variety and lightweight

classification network will be studied with a high accuracy rate and rapid speed in future
work to lay the foundation for the research and development of multi-variety classification
equipment.
The actual situation for the classification of koi varieties may be complex. Therefore,
a multi-objective situation may occur, and some factors (such as posture change of koi,
object occlusion, and illumination change) affect the classification accuracy of koi varieties.
Our future work will focus on solving the problems of classifying koi varieties in complex
situations.

5. Conclusions
Koi variety classification was studied to solve the problems caused by the high simi-
larity among some varieties. In this study, a dataset including thirteen kinds of koi was
established, and a koi variety classification network KRS-Net was proposed based on resid-
ual network and SVM. Compared with five other mainstream networks, the performance
superiority of the proposed KRS-Net was proven. this study provides a new solution for
the classification of koi varieties, which can be extended to breeding, aquaculture, grade
sorting, and other marine fields.

Author Contributions: Conceptualization, J.L.; methodology, J.L.; software, Y.Z.; validation, J.L. and
L.D.; formal analysis, F.W.; investigation, Y.Z. and Q.L.; resources, F.W.; data curation, F.W.; writing—
original draft preparation, Y.Z.; writing—review and editing, J.L.; visualization, W.X.; supervision,
J.L. and L.D.; project administration, J.L.; funding acquisition, J.L. All authors have read and agreed
to the published version of the manuscript.
Funding: This research work was financially supported by the project of the National Natural
Science Foundation of China (32073029), the key project of the Shandong Provincial Natural Science
Foundation (ZR2020KC027), the postgraduate education quality improvement project of Shandong
Province (SDYJG19134), the project of the China Scholarship Council (201908370048), and the Open
Program of Key Laboratory of Cultivation and High-value Utilization of Marine Organisms in Fujian
Province (2021fjscq08).
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: With the consent of the authors, the data to this article can be found
online at: http://33596lk211.qicp.vip/dataweb/index.html, accessed on 21 May 2022.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Pinkey, S. Study of the fresh water fish diversity of Koshi river of Nepal. Int. J. Fauna Biol. Stud. 2016, 3, 78–81.
2. Nuwansi, K.; Verma, A.; Chandrakant, M.; Prabhath, G.; Peter, R. Optimization of stocking density of koi carp (Cyprinus carpio
var. koi) with gotukola (Centella asiatica) in an aquaponic system using phytoremediated aquaculture wastewater. Aquaculture
2021, 532, 735993. [CrossRef]
3. Wang, K.L.; Chen, S.N.; Huo, H.J.; Nie, P. Identification and expression analysis of sixteen Toll-like receptor genes, TLR1, TLR2a,
TLR2b, TLR3, TLR5M, TLR5S, TLR7-9, TLR13a-c, TLR14, TLR21-23 in mandarin fish Siniperca chuatsi. Dev. Comp. Immunol. 2021,
121, 104100. [CrossRef]
4. De Kock, S.; Gomelsky, B. Japanese Ornamental Koi Carp: Origin, Variation and Genetics; Biology and Ecology of Carp; Informa UK
Limited: Boca Raton, FL, USA, 2015; pp. 27–53.
5. Sun, X.; Chang, Y.; Ye, Y.; Ma, Z.; Liang, Y.; Li, T.; Jiang, N.; Xing, W.; Luo, L. The effect of dietary pigments on the coloration of
Japanese ornamental carp (koi, Cyprinus carpio L.). Aquaculture 2012, 342, 62–68. [CrossRef]
6. Bairwa, M.K.; Saharan, N.; Rawat, K.D.; Tiwari, V.K.; Prasad, K.P. Effect of LED light spectra on reproductive performance of Koi
carp (Cyprinus carpio). Indian J. Anim. Res. 2017, 51, 1012–1018. [CrossRef]
7. Xie, Z.; Wang, D.; Jiang, S.; Peng, C.; Wang, Q.; Huang, C.; Li, S.; Lin, H.; Zhang, Y. Chromosome-Level Genome Assembly
and Transcriptome Comparison Analysis of Cephalopholis sonnerati and Its Related Grouper Species. Biology 2022, 11, 1053.
[CrossRef]
8. Nica, A.; Mogodan, A.; Simionov, I.-A.; Petrea, S.-M.; Cristea, V. The influence of stocking density on growth performance of
juvenile Japanese ornamental carp (Koi, Cyprinus carpio L.). Sci. Pap. Ser. D Anim. Sci. 2020, 63, 483–488.
Biology 2022, 11, 1727 15 of 16

9. Kim, J.-I.; Baek, J.-W.; Kim, C.-B. Image Classification of Amazon Parrots by Deep Learning: A Potentially Useful Tool for Wildlife
Conservation. Biology 2022, 11, 1303. [CrossRef]
10. Tian, X.; Pang, X.; Wang, L.; Li, M.; Dong, C.; Ma, X.; Wang, L.; Song, D.; Feng, J.; Xu, P. Dynamic regulation of mRNA and miRNA
associated with the developmental stages of skin pigmentation in Japanese ornamental carp. Gene 2018, 666, 32–43. [CrossRef]
11. Peng, F.; Chen, K.; Zhong, W. Classification and appreciation of three species of koi. Sci. Fish Farming 2018, 8, 82–83.
12. Song, S.; Duan, P. Koi and its variety classification. Shandong Fish. 2009, 26, 53–54.
13. Garland, J.; Hu, M.; Kesha, K.; Glenn, C.; Duffy, M.; Morrow, P.; Stables, S.; Ondruschka, B.; Da Broi, U.; Tse, R. An overview of
artificial intelligence/deep learning. Pathology 2021, 53, S6. [CrossRef]
14. Talib, M.A.; Majzoub, S.; Nasir, Q.; Jamal, D. A systematic literature review on hardware implementation of artificial intelligence
algorithms. J. Supercomput. 2021, 77, 1897–1938. [CrossRef]
15. Liu, Y.; Yang, C.; Gao, Z.; Yao, Y. Ensemble deep kernel learning with application to quality prediction in industrial polymerization
processes. Chemom. Intell. Lab. Syst. 2018, 174, 15–21. [CrossRef]
16. Wang, T.; Zhao, Y.; Sun, Y.; Yang, R.; Han, Z.; Li, J. Recognition approach based on data-balanced faster R CNN for winter jujube
with different levels of maturity. Trans. Chin. Soc. Agric. Mach. 2020, 51, 457–463.
17. Xu, W.; Zhao, L.; Li, J.; Shang, S.; Ding, X.; Wang, T. Detection and classification of tea buds based on deep learning. Comput.
Electron. Agric. 2022, 192, 106547. [CrossRef]
18. Li, J.; Xu, C.; Jiang, L.; Xiao, Y.; Deng, L.; Han, Z. Detection and analysis of behavior trajectory for sea cucumbers based on deep
learning. IEEE Access 2019, 8, 18832–18840. [CrossRef]
19. Xu, W.; Zhu, Z.; Ge, F.; Han, Z.; Li, J. Analysis of behavior trajectory based on deep learning in ammonia environment for fish.
Sensors 2020, 20, 4425. [CrossRef] [PubMed]
20. Li, J.; Xu, W.; Deng, L.; Xiao, Y.; Han, Z.; Zheng, H. Deep Learning for Visual Recognition and Detection of Aquatic Animals: A
Review. Rev. Aquac. 2022, 1–25. [CrossRef]
21. Xu, Y.; Shen, H. Review of Research on Biomedical Image Processing Based on Pattern Recognition. J. Electron. Inf. Technol. 2020,
42, 201–213.
22. Sarica, A.; Vaccaro, M.G.; Quattrone, A.; Quattrone, A. A Novel Approach for Cognitive Clustering of Parkinsonisms through
Affinity Propagation. Algorithms 2021, 14, 49. [CrossRef]
23. Liu, W.; Wang, Z.; Zeng, N.; Alsaadi, F.E.; Liu, X. A PSO-based deep learning approach to classifying patients from emergency
departments. Int. J. Mach. Learn. Cybern. 2021, 12, 1939–1948. [CrossRef]
24. Han, Z.; Wan, J.; Deng, L.; Liu, K. Oil Adulteration identification by hyperspectral imaging using QHM and ICA. PLoS ONE 2016,
11, e0146547. [CrossRef] [PubMed]
25. Liu, W.; Wang, Z.; Liu, X.; Zeng, N.; Liu, Y.; Alsaadi, F.E. A survey of deep neural network architectures and their applications.
Neurocomputing 2017, 234, 11–26. [CrossRef]
26. Eerola, T.; Kraft, K.; Grönberg, O.; Lensu, L.; Suikkanen, S.; Seppälä, J.; Tamminen, T.; Kälviäinen, H.; Haario, H. Towards
operational phytoplankton recognition with automated high-throughput imaging and compact convolutional neural networks.
Ocean. Sci. Discuss. 2020, 62, 1–20.
27. Zhu, D.; Qi, R.; Hu, P.; Su, Q.; Qin, X.; Li, Z. YOLO-Rip: A modified lightweight network for Rip Currents detection. Front. Mar.
Sci. 2022, 9, 930478. [CrossRef]
28. Lu, Y.; Tung, C.; Kuo, Y. Identifying the species of harvested tuna and billfish using deep convolutional neural networks. ICES J.
Mar. Sci. 2020, 77, 1318–1329. [CrossRef]
29. Rauf, H.T.; Lali, M.I.U.; Zahoor, S.; Shah, S.Z.H.; Rehman, A.U.; Bukhari, S.A.C. Visual features based automated identification of
fish species using deep convolutional neural networks. Comput. Electron. Agric. 2019, 167, 105075. [CrossRef]
30. Knausgård, K.M.; Wiklund, A.; Sørdalen, T.K.; Halvorsen, K.T.; Kleiven, A.R.; Jiao, L.; Goodwin, M. Temperate fish detection and
classification: A deep learning based approach. Appl. Intell. 2022, 52, 6988–7001. [CrossRef]
31. Ju, Z.; Xue, Y. Fish species recognition using an improved AlexNet model. Optik 2020, 223, 165499. [CrossRef]
32. Hridayami, P.; Putra, I.K.G.D.; Wibawa, K.S. Fish species recognition using VGG16 deep convolutional neural network. J. Comput.
Sci. Eng. 2019, 13, 124–130. [CrossRef]
33. Huang, X.; Chen, W.; Yang, W. Improved Algorithm Based on the Deep Integration of Googlenet and Residual Neural Network; IOP
Publishing: Bristol, UK, 2021; p. 012069.
34. Xu, X.; Li, W.; Duan, Q. Transfer learning and SE-ResNet152 networks-based for small-scale unbalanced fish species identification.
Comput. Electron. Agric. 2021, 180, 105878. [CrossRef]
35. Zhang, K.; Guo, Y.; Wang, X.; Yuan, J.; Ding, Q. Multiple feature reweight densenet for image classification. IEEE Access 2019, 7,
9872–9880. [CrossRef]
36. Gao, S.; Dai, Y.; Li, Y.; Jiang, Y.; Liu, Y. Augmented flame image soft sensor for combustion oxygen content prediction. Meas. Sci.
Technol. 2022, 34, 015401. [CrossRef]
37. Liu, K.; Li, Y.; Yang, J.; Liu, Y.; Yao, Y. Generative principal component thermography for enhanced defect detection and analysis.
IEEE Trans. Instrum. Meas. 2020, 69, 8261–8269. [CrossRef]
38. Zhu, H.; Yang, L.; Fei, J.; Zhao, L.; Han, Z. Recognition of carrot appearance quality based on deep feature and support vector
machine. Comput. Electron. Agric. 2021, 186, 106185. [CrossRef]
Biology 2022, 11, 1727 16 of 16

39. Hemke, R.; Buckless, C.G.; Tsao, A.; Wang, B.; Torriani, M. Deep learning for automated segmentation of pelvic muscles, fat, and
bone from CT studies for body composition assessment. Skelet. Radiol. 2020, 49, 387–395. [CrossRef]
40. Selvaraju, R.R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; Batra, D. Grad-CAM: Visual Explanations from Deep Networks
via Gradient-based Localization. Int. J. Comput. Vis. 2020, 128, 336–359. [CrossRef]
41. Taheri-Garavand, A.; Ahmadi, H.; Omid, M.; Mohtasebi, S.S.; Carlomagno, G.M. An intelligent approach for cooling radiator
fault diagnosis based on infrared thermal image processing technique. Appl. Therm. Eng. 2015, 87, 434–443. [CrossRef]
42. Zhang, Z.; Tang, Y.-G.; Yang, K. A two-stage restoration of distorted underwater images using compressive sensing and image
registration. Adv. Manuf. 2021, 9, 273–285. [CrossRef]
43. Burçak, K.C.; Baykan, Ö.K.; Uğuz, H. A new deep convolutional neural network model for classifying breast cancer histopatho-
logical images and the hyperparameter optimisation of the proposed model. J. Supercomput. 2021, 77, 973–989. [CrossRef]
44. Goyal, P.; Dollár, P.; Girshick, R.; Noordhuis, P.; Wesolowski, L.; Kyrola, A.; Tulloch, A.; Jia, Y.; He, K. Accurate, large minibatch
sgd: Training imagenet in 1 hour. arXiv 2017, arXiv:1706.02677.