https://doi.org/10.31449/inf.v48i9.5929 Informatica 48 (2024) 143–154 143 
Basketball Video Image Segmentation Using Neutrosophic Fuzzy 
C-means Clustering Algorithm 
Chao Hong 
Ministry of Sports, Xiamen Institute of Technology, Xiamen 361021, China 
E-mail: hc_0727@163.com 
Keywords: video segmentation, fuzzy C-means clustering, neutrosophic fuzzy C-means clustering, basketball video 
images 
Received: March 18, 2024 
Basketball video image segmentation is important in image processing and computer vision, which is 
significant for improving image quality and enhancing visual effects. However, traditional image 
segmentation algorithms still have challenges in dealing with noise and complex backgrounds. 
Research on basketball video image segmentation based on Neutrosophic fuzzy C-means clustering 
algorithm was conducted in this paper. Firstly, the video image segmentation algorithm was studied 
and analyzed. Secondly, the computational time was reduced to get better segmentation results by fuzzy 
C-mean clustering. The algorithm was carried out for basketball video image segmentation, which was 
compared and analyzed with the traditional segmentation algorithm. Results showed that the peak 
SNR values were 14.96, 14.81, and 14.57 in pretzel noise environment. The peak SNR results were 
13.97, 12.87, and 12.06 in Gaussian noise environment. The algorithm has a significant advantage in 
both image segmentation performance. It improves the image quality and visual effect, which is an 
important reference for future image analysis and processing. 
Povzetek: Algoritem za segmentacijo video slik v košarki uporablja gručenje Neutrosophic fuzzy 
C-means, kar izboljša kakovost slike in vizualne učinke ter zmanjša čas izračuna v primerjavi s 
tradicionalnimi metodami.
1 Introduction 
Basketball is a fast-paced and highly competitive sport 
with a rich variety of dynamic images and emotional 
changes. The accurate video image segmentation is a 
challenge [1]. Traditional video image segmentation 
methods can hardly deal with the complex and dynamic 
scenes in basketball games [2]. The Fuzzy C-means 
Clustering (ZFC) algorithm is a classical clustering 
segmentation method that can manage images containing 
noise and fuzzy edges [3]. However, the classical 
clustering segmentation method depends on the initial 
value and may fall into local optimal solutions [4]. An 
important research topic in video image segmentation 
technology is to improve the stability and accuracy of the 
ZFC algorithm [5]. In this study, a basketball video image 
segmentation method using Neutrosophic Fuzzy C-means 
Clustering (ZZFC) algorithm is proposed. This method 
aims to optimize the initial value selection and adopt 
intelligent optimization strategies to improve the stability 
and accuracy of segmentation. In this study, ZZFC is 
innovatively applied to basketball video image 
segmentation. Meanwhile, two strategies, center updating 
and intelligent optimization, are adopted to effectively 
improve the stability and accuracy of clustering. This 
study is important in promoting basketball video image 
segmentation technology. Therefore, accurate and 
high-quality data support and important value can be 
supplied for subsequent game analysis and athlete 
performance evaluation. 
The research includes five parts. The first part 
describes how video image segmentation technology 
promotes computer and information science with the 
rapid development of information technology. The 
second part is the related works, which elaborate on the 
current research and implementation of video image 
segmentation algorithms by numerous scholars. The third 
part is the video image segmentation technology using 
ZZFC. The first section focuses on video image 
segmentation algorithms, while the second section 
focuses on the video image segmentation algorithms 
under ZZFC. In the fourth part, the proposed algorithm is 
tested. The fifth part is a summary and outlook of this 
study. 
2 Related works 
In the current research, basketball video image 
segmentation is a hot topic. However, traditional image 
segmentation methods often do not achieve ideal results 
due to the complexity and diversity. Many scholars have 
studied video image segmentation techniques. Fiaz et al. 
proposed a network architecture called 4G video object 
segmentation. This architecture improved the video 
object segmentation by encoding video context to address 
the background interference and object appearance 
changes. The guided transfer embedding module was 

144   Informatica 48 (2024) 144–154                                                                 X. Li
 
et al. 
used to keep long-term semantic information in this 
architecture, and the global instance matching module 
generated similar mappings. The experimental results 
showed that 4G video object segmentation performed 
well among 40 existing methods [6]. Cao et al. proposed 
a segmentation scheme with two branches. The attention 
mechanism was first used to highlight features related to 
objects. Meanwhile, a designed attention residual 
convolution was used to capture the long and short-term 
temporal information of objects under redundant video 
frame interference. The segmented target was obtained 
using global thresholding and noise region removal 
methods based on the fusion results of the two branches. 
The experimental results demonstrated the competitive 
performance of the scheme [7]. Zhang et al. proposed a 
new method for monitoring video motion segmentation 
based on a progressive spatiotemporal tunnel flow model. 
A cyclic progressive spatiotemporal tunnel was 
established by gradually sampling pixels, and then it was 
unfolded to form a spatiotemporal tunnel unfolding map. 
Finally, the monitoring video motion segments were 
segmented based on this model. Results showed that this 
method was superior to existing methods in time 
consumption [8]. Zhu et al. proposed an algorithm for 
non-specific sperm aggregates. This algorithm created a 
grid model that was proportional to the sperm head size 
to segment non-specific aggregation areas. Meanwhile, 
multi-scale edge functions and new energy functions 
were designed to achieve sperm head segmentation. The 
research results indicated that this method achieved 
precise segmentation of sperm, outperforming the level 
set methods. The sperm concentration and vitality were 
calculated in real-time during sperm tracking [9]. Wang 
et al. proposed a feature reconstruction-based method for 
human portrait video segmentation. This method 
introduced a soft communication network that promoted 
feature reconstruction in an unsupervised manner by 
gently allocating the displacement probability of each 
pixel. More reliable segmentation results could be 
obtained by mining reconstructed features. The research 
results indicated the effectiveness of the method [10]. 
The ZZFC algorithm is a novel method for 
basketball video image segmentation, showing better 
advantages. Many scholars have conducted extensive 
research on it. Yang et al. proposed a fuzzy C-means 
algorithm that simultaneously handled noise and 
uncertainty issues in MRI images of human brain. Noise 
robust intuitionistic fuzzy sets were introduced to better 
handle images with noise. The research results indicated 
that the algorithm demonstrated effectiveness and 
superiority in many MRI experiments of human brain 
[11]. Z Dong et al. proposed a cold damage area 
segmentation algorithm using improved k-means 
clustering for detecting tomato frost damage. The 
research results indicated that the algorithm performed 
well in segmenting cold damaged areas in fluorescence 
imaging with yellow cold spots. The average matching 
rate and average error rate were better than the traditional 
methods [12]. K G Dhal et al. proposed a fuzzy clustering 
technique that combined rough set-based global, random 
attraction, and local search strategies to deal with image 
segmentation. The research results indicated that this 
technology outperformed various natural inspired 
algorithms in accuracy and segmentation output quality 
[13]. Bas et al. proposed a fuzzy inference system using 
Gustafson Kessel clustering algorithm. A fuzzy 
regression function method was used instead of rule base 
operation. This new method recognized ellipsoidal 
clustering without traditional limitations. The research 
results indicated that the fuzzy inference system using 
this clustering algorithm significantly improved 
predictive performance [14]. Li et al. proposed a spectral 
clustering algorithm using fuzzy similarity. The fuzzy 
similarity measurement was used to obtain the similarity 
between data points and anchor points, and then the 
similarity matrix was obtained. Results showed that the 
proposed algorithm had better classification performance 
for hyperspectral remote sensing images, with a 2% 
increase in kappa coefficient compared to traditional 
algorithms [15]. The summary of relevant research is 
shown in Table 1. 
 
Table1: Summary table of related research 
Method Data set Key findings Limitations 
M Fiaz et al. 
DA VIS2016, DA VIS2017 and 
YouTube VOS 
Has excellent image 
segmentation performance 
Low efficiency 
Y Cao et al. DA VIS2016 and CDnet2014 
Performed well in long 
video segmentation tasks 
It is difficult to segment 
small object targets from 
videos with occlusion 
Y Zhang et al. 8 real monitoring scenarios Less time consumption 
Lack of motion 
classification for motion 
segments 
R Zhu et al. Real sperm samples 
Accurate segmentation of 
non-specific aggregation 
regions has been achieved 
High demand for image 
quality 
Y Wang et al. DA VIS2016 and CDnet2014 
Suitable for scenarios with 
significant posture changes 
Low efficiency 

Basketball Video Image Segmentation Using Neutrosophic Fuzzy… Informatica 48 (2024) 145–154 145 
and severe occlusion 
Z Yang et al. Brain Web 
Combined with local 
spatial information, less 
affected by noise 
Computational complexity 
Z Dong et al. 
Fluorescence images of tomato 
seedling leaves collected 
The average matching rate 
of the image is relatively 
high 
Poor processing effect for 
complex scenes 
K Li et al. 
Synthetic data and real word 
hyperspectral imaging dataset 
Simple calculation, suitable 
for large-scale data 
classification such as 
hyperspectral remote 
sensing images 
Poor stability 
 
In summary, the existing advanced basketball video 
image segmentation methods have made great progress. 
However, there are still problems such as high 
computational complexity, high requirements for image 
quality, poor performance in complex scene processing, 
and poor stability. Therefore, this study proposes a 
basketball video image segmentation method based on 
the ZZFC algorithm. Specifically, the image 
segmentation method using the ZZFC introduces the 
Neutrosophic fuzzy theory to better handle uncertainty in 
images. Then the Neutrosophic fuzzy theory is combined 
with the C-means clustering algorithm to achieve more 
accurate image segmentation. As a result, image 
segmentation and noise can be handled effectively to 
preserve important features of the image. However, this 
method still has certain shortcomings, such as high 
computational complexity and difficulties in large-scale 
image processing. Therefore, the proposed methods still 
need to be improved and optimized to provide important 
directions and basis for future research. 
3 Video image segmentation 
algorithm using neutrosophic fuzzy 
C-means clustering algorithm 
Video image segmentation algorithms focus on image 
segmentation in video streams to improve image 
processing efficiency and accuracy. On this basis, the  
 
 
ZZFC algorithm is introduced to further improve the 
accuracy of image segmentation. The ZZFC algorithm 
can significantly reduce the impact of errors in image 
segmentation with the high adaptability to fuzziness and 
noise. This algorithm not only promotes the development 
of efficient video processing technology, but also 
provides new theoretical and application foundations for 
innovative exploration of image segmentation technology. 
It is hoped to provide new evidence and explore new 
perspectives for researchers in deep learning and machine 
vision. 
3.1 Basketball video processing based on 
image segmentation algorithm 
Video image segmentation, as a key technology in 
computer vision, has expanded the research space in 
multiple scientific fields such as modern image 
processing, machine learning, and artificial intelligence 
[16]. This technology analyzes the pixel population 
within a video frame to determine the object or scene it 
belongs to, thereby achieving precise analysis and 
understanding of video content. Video image 
segmentation algorithms rapidly evolve with the 
development of the combination of deep learning and 
image segmentation. The development of this method 
helps to improve video processing efficiency and 
accuracy. The video image segmentation algorithm is 
shown in Figure 1. 
 
Start
Set algorithm parameters and select the 
maximum inter class variance method as the 
objective optimization function
Random initialization
Evaluate and record the optimal individual 
based on the objective function value
Is the 
algorithm over
Use pattern search algorithm to re search 
local video individuals
Generate new signals iteratively through the 
strength mechanism of video image signals
Output optimal individual
End
Yes
No
No
 
Figure 1: Video image segmentation algorithm 
 
The algorithm consists of three main parts: 
preprocessing, segmentation, and post-processing [17]. In 
the preprocessing, video images are optimized through 
filtering, noise suppression, and other methods to 

146   Informatica 48 (2024) 146–154                                                                 X. Li
 
et al. 
improve the accuracy of subsequent segmentation. The 
segmentation is the core of the algorithm, which 
distinguishes foreground objects and background scenes 
in the image. The segmentation usually relies on a 
generator network to achieve pixel level fine 
segmentation. In the post-processing, operations such as 
edge smoothing and hole filling are performed to improve 
the visual representation of the segmentation results. The 
frame difference method extracts the image background 
of the video motion target detection system, as calculated 
in equation (1). 
 
( , ) ( , , ) ( , , 1)
t
I x y I x y t I x y t  = − −
 (1) 
In equation (1), ( , , ) I x y t represents the grayscale 
of the pixel with coordinate ( , ) xy in frame t of the 
video. ( , , 1) I x y t − represents the grayscale of the pixel 
with ( , ) xy in 1 t − of the video. ( , )
t
I x y  is the 
difference between the grayscale of ( , , ) I x y t and 
( , , 1) I x y t − . When the algorithm detects that the 
foreground target has voids inside the object and there is 
stretching distortion in the motion direction of the object, 
its mathematical expression is shown in equation (2). 
 
1, ( , )
( , )
0, ( , )
t
t
t
I x y
D x y
I x y


  

=




 (2) 
In equation (2), 

 represents the threshold for 
restricting ( , )
t
I x y  , and 
t
D represents the background 
for obtaining frame t . The formula for calculating the 
difference between the three frames is shown in equation 
(3) for three consecutive frames of video images. 
( , ) ( , , ) ( , , 1) ( , , ) ( , , 1)
t
I x y I x y t I x y t I x y t I x y t  = − −  − −
 (3) 
In equation (3), when the ( , )
t
I x y  value in the 
area of frame t image is small, the possibility of its 
background is greater. When the ( , )
t
I x y  value in the 
field is high, foreground moving targets likelihood is 
higher. The calculation formula for the binary image 
( , )
t
W x y of its three-frame difference is shown in 
equation (4). 
 
1, ( , )
( , )
0, ( , )
t
t
t
I x y
W x y
I x y


  

=




 (4) 
In equation (4), 

 represents the threshold 
introduced to restrict ( , )
t
I x y  . In image segmentation, 
noise that affects image quality is easily generated, so it is 
necessary to preprocess video images. The video 
preprocessing combining reinforcement learning methods 
is shown in Figure 2.
  
Original video 
sequence
Background 
subtraction
morphology filter
Hyperpixel 
segmentation
Preprocessed video 
sequence
 
Figure 2: Video preprocessing process combining reinforcement learning methods 
 
Video preprocessing is a key step in video image 
segmentation algorithms, aiming at optimizing the input 
video data to improve segmentation performance [18]. 
The preprocessing process first involves the collection 
and decoding of video frames, converting the raw data 
into a form that can be processed by the computer. The 
preprocessing processes are together constituted to 
provide higher quality input data for subsequent video 
image segmentation. The low-pass filtering method is 
used to convolve the impulse response functions of 
images and filters in the spatial domain. The 
mathematical expression of the convolution is shown in 
equation (5). 
( , ) ( , ) ( 1, 1)
xy
g u v f x y H u x v y = − + − +

 (5) 
In equation (5), ( , ) f x y represents the image. 
( , ) H x y represents the impulse response function of the 
filter. The motion video segmentation algorithm using 
global motion constraints and Markov random fields is a 
complex and effective image processing method. The 
video segmentation algorithm is shown in Figure 3. 
 

Basketball Video Image Segmentation Using Neutrosophic Fuzzy… Informatica 48 (2024) 147–154 147 
Original motion 
video image
Multiple image 
feature sequences
Object prediction 
sequence
Transformer encoder 
decoder
Deep learning 
network
 
Figure 3: A video segmentation algorithm using global motion constraints and markov random fields 
 
First, the video segmentation algorithm using global 
motion constraints and Markov random fields 
preprocesses the video frames using global motion 
constraints. Then local motion information of foreground 
objects is obtained by estimating and removing global 
motion. Next, Markov random field model is established, 
which combines the spatial relationships between pixels 
and time series information to achieve video frame 
segmentation. The calculation formula for setting the 
coordinate points ( , ) xy and 
''
( , ) xy corresponding to 
the current frame and the previous frame is shown in 
equation (6). 
'' 0 1 2 3 4 5
6 7 6 7
,
11
m x m y m m x m y m
xy
m x m y m x m y
+ + + +
==
+ + + +
 (6) 
In equation (6), 
m
 represents global motion 
parameters. 
t
m represents the motion model parameters 
calculated at frame t . The mathematical expression is 
shown in equation (7). 
( , , ) ( , , ) ( , , )
res
t
MV x y t MV x y t MV x y m =− (7) 
In equation (7), ( , , )
res
MV x y t represents the final  
 
 
MV calculation result. The calculation formula for 
clustering the motion vectors is shown in equation (8). 
 
/
res
cent j j j
jj
MV W MV W =

 (8) 
In equation (8), 
res
j
MV represents the motion 
vector of each individual, and 
exp( )
j MAXj
WD =−
 
represents the weights in the domain. The video image 
segmentation algorithm accurately segments complex 
basketball game scenes to further provide accurate 
analysis of motion trajectories, speeds, and behavioral 
patterns. This algorithm plays a key role in promoting 
basketball training, game strategy development, and 
viewing experience. 
3.2 Video image segmentation using 
neutrosophic fuzzy C-means clustering 
algorithm 
The quality of video image segmentation directly affects 
the application of video [19]. The traditional method is 
prone to problems such as initial value selection, noise 
influence, and extreme value sensitivity when processing 
complex video images, as shown in Figure 4. These 
challenges can be solved by introducing fuzzy logic and 
Neutrosophic theory. 
 
Initialize sample membership matrix and cluster 
center
Number of initialization iterations and amount 
of cluster center transformation
Is the change in cluster center less 
than
Has the maximum number of 
iterations been reached
Update membership matrix
Update Cluster Center
Update the number of iterations and calculate 
the change in clustering centers
At the end of the iteration, 
determine the final category 
based on membership 
hardening
Yes
Yes
No
No
 
Figure 4: Fuzzy C-means clustering algorithm 

148   Informatica 48 (2024) 148–154                                                                 X. Li
 
et al. 
 
The ZFC algorithm allows data points to belong to 
multiple cluster centers by hard C-means clustering. The 
membership degree of each cluster is a continuous value 
between 0 and 1. The ZFC algorithm adjusts the center 
and membership alternately until the end condition is met 
or the preset number of iterations is reached. The fuzzy 
clustering method is applied in image segmentation 
problems. Meanwhile, the mean clustering method with 
parameter z is introduced. The mathematical 
expression of the objective function is shown in equation 
(9). 
 
2
11
( , ) ( , )
Nc
m
ij i j
ij
J U V u d x v
==
=
 (9) 
In equation (9), 
m
 represents the fuzzy parameter, 
with a value of 2. The membership of 
ij
u
 in 
ij
Uu  =

 
represents the degree to which each data point 
n
x 
belongs to different categories, with a value range of 
  0,1 . The total membership degree of each sample set is 
1, so the range of U is shown in equation (10). 
 
11
0,1 1, 0 ,
cN
ij ij ij
ji
U u u kand u N i
==


  =    




(10) 
In equation (10), 
ij
u
 represents updating the cluster 
center. The ZZFC algorithm partitions the elements in the 
dataset into various clusters, while allowing them to 
belong to multiple clusters [20]. The ZZFC algorithm is 
shown in Figure 5. 
 
INitialization
Initialization parameters
Calculate the center vector of formula 
K step
Calculate          using the calculation method based on the 
maximum and second maximum values of 
Update T 
(k)
 to T 
(k+1)
, I 
(k)
 to I 
(k+1)
, T 
(0)
 to I 
(0) 
using 
equations
If                         , stop, otherwise 
return to step three
Set each data to the maximum                      value in the class, and the 
result for class        is      
(0)
T
(0)
I
(0)
F
C m  
1
w
2
w
3
w
() k
c
( 1) ( ) kk
TT 
+
−
T
  ,, TM T I F =
() xi
1,2, , 2
arg max ( )
ij
jc
k TM
=+
=
,max i C
 
Figure 5: Neutrosophic fuzzy C-means clustering algorithm 
 
The membership in this algorithm is calculated by 
fuzzy logic, and the result can better reflect the fuzziness 
between data. The algorithm iteratively updates the 
membership of cluster centers and elements until the 
predetermined stopping conditions are reached. The 
objective function of the two closest determined clusters 
is simplified by only considering the maximum and 
second largest membership values. Meanwhile, the 
computational cost is reduced, but the clustering accuracy 
will not be reduced. The expression of its objective 
function is shown in equation (11). 
2 2
2
2 ,max 3
1 1 1 1
min ( , , , ) ( ) ( ) ( )
mm
n c u n
m
m I ij i j i i i i
i j i i
J U I F V w u x v w I x v w F 
= = = =
= − + − +
    (11) 
In equation (11), 
m
 is a constant. 
i
P and 
e
P 
represent the number of clusters with the highest and 
second highest values. The datasets for boundary areas 
and noise, 
0
ij
T 
, 
i
I , and 1
i
F  , the following 
equations are satisfied, as shown in equation (12). 
 
1
1
C
ij i i
j
T I F
=
+ + =
 (12) 
In equation (12), 
i
T , 
ij
T
, 
i
I , and 
i
F represent the 
membership that belongs to the determined cluster. The 
ZZFC video segmentation algorithm is an optimization 
method using fuzzy logic and C-means clustering, 
focusing on efficient segmentation of video content to 
obtain finer content structures. The ZZFC video 
segmentation algorithm is shown in Figure 6. 
 

Basketball Video Image Segmentation Using Neutrosophic Fuzzy… Informatica 48 (2024) 149–154 149 
Size M × N's grayscale image G input
Using the mountain climbing method to select the initialization cluster center      , 
set the algorithm iteration convergence control error term e=0.0001, and the 
number of iterations t=0
(0)
j
v
Determining kernel function parameters in intelligent fuzzy clustering in kernel 
space based on statistical information of grayscale images
According to the calculation of membership degree and cluster center 
corresponding to the FCM algorithm used by       , the initialization cluster center 
of intelligent fuzzy C-means clustering in the kernel space is obtained
Calculate membership based on intelligent fuzzy clustering image segmentation 
algorithm in kernel space
Does it meet 
Further improve the anti noise ability of the clustering segmentation algorithm 
based on the gray level membership corresponding to the proposed algorithm
Stopping Intelligent Fuzzy C-
means Clustering in Kernel Space
Using             to classify pixel values based on 
the principle of maximum membership
Output segmented image       
(0)
j
v
( 1) t
j
v
+
,, x y j
u

'
G
Yes
No
 
Figure 6: Neutrosophic fuzzy C-means clustering video segmentation 
 
This algorithm compensates for the shortcomings of 
traditional ZFC algorithms in handling image noise and 
detail complexity by introducing Neutrosophic fuzzy 
logic. Firstly, the video frames are converted into a fuzzy 
set, and then C-means clustering processes the fuzzy set, 
forming multiple non-overlapping clusters. After 
processing the image information, the loss of the original 
image quality is calculated using the mean square error 
method, as shown in equation (13). 
 
2
11
00
1
( , ) ( , )
mn
ij
MSE I i j K i j
mn
−−
==
=−

 (13) 
In equation (13), I and K represent two 
monochromatic graphs. The mathematical expression for 
its peak signal-to-noise ratio is shown in equation (14). 
2
10 10
10 log ( ) 20 log ( )
J I
MAX MAX
PSNR
MSE MSE
=  =  (14) 
In equation (14), 
I
MAX represents the maximum 
value of image points color. The algorithm can achieve 
efficient recognition of inter frame coherence and video 
content while ensuring segmentation accuracy through 
iterative optimization. 
 
 
 
4 Test analysis of video image 
segmentation algorithm using 
neutrosophic fuzzy C-means 
clustering algorithm 
The study used the Potsdam dataset to verify the 
effectiveness and superiority of the segmentation 
performance of the ZZFC algorithm with the proposed 
local information constraints for segmentation of images 
[21]. The ZFC, FCM, and ZZFC algorithms were selected 
to have good noise immunity. The maximum iteration 
was set to 300, the weight of the criterion function was 
set to 1.5, the clusters were set to 8, the fuzzy factor was 
set to 2, the error was set to 0.001, and the classifications 
were set to 2. The hardware environment for testing and 
analysis included a computer with Intel Core i7-9700K 
processor with 32GB of RAM and 1TB SSD as hard disk. 
The experimental running environment windows XP with 
MATLAB 7.0 was chosen as the development platform. 
The programming environment was Python 3.8. OpenCV 
4.5.1 was used for video image processing and 
Scikit-learn 0.24.1 library was used to implement the 
ZZFC algorithm. In addition, the NumPy 1.20.1 library 
was used for high performance scientific computing and 
data analysis, and the Matplotlib 3.3.4 library was used 
for image display and result plotting. The segmentation 
results of different algorithms for the figure and cotton 
images are shown in Figure 7. 

150   Informatica 48 (2024) 150–154                                                                 X. Li
 
et al. 
 
(b) Segmentation Results of Image Cotton
(a) Segmentation Results of Image Characters
Original image ZFC FCM ZFCM
Original image
ZFC
FCM ZFCM ZFC
 
Figure 7: Segmentation results of different algorithms for character and cotton images 
 
In Figure 7, the superior performance of the ZZFC 
algorithm in image detail segmentation was demonstrated, 
and its segmentation accuracy was also significantly 
improved. The ZZFC algorithm showed a more precise 
segmentation effect, with a significant reduction in 
misclassification in the segmentation experiment of 
character images in Figure 7 (a). The ZZFC algorithm 
had significantly overall evaluation performance better 
than ZFC and FCM algorithms. In Figure 7 (b), the ZZFC 
algorithm significantly improved the segmentation 
accuracy in details such as windows and stairs, with 
strong detail expressiveness. The performance of the 
ZZFC algorithm clearly surpassed the other two 
segmentation algorithms. The advantages of the ZZFC 
algorithm in image segmentation provided new ideas and 
methods for further research on image segmentation 
technology. The PSNR results of different algorithms are 
shown in Figure 8. 
 
ZFC
FCM
ZFCM
35% 55% 45%
0
10
Noise standard deviation
Average PSNR (dB)
20
(a) Salt and pepper noise
Noise standard deviation
Average PSNR (dB)
(b) True noise
35% 55% 45%
15
5
0
10
20
15
5
ZFC
FCM
ZFCM
 
Figure 8: PSNR results of different algorithms under spicy salt noise interference 
 
Figure 8 (a) shows the average PSNR performance 
of each algorithm. The PSNR values achieved by the 
ZZFC algorithm all showed the best performance when 
the noise standard deviation was set to 35, 45, and 55. 
The average PSNR improvement ratio of ZZFC 
compared to the FCM algorithm was 13.45%, 11.54%, 
and 9.19%, respectively, when the noise standard 
deviation was 35, 45, and 55. Similarly, the average 
PSNR improvement of ZZFC compared to ZFC 
algorithm was 21.94%, 21.74%, and 23.68%, respectively, 
when the noise standard deviation was 35, 45, and 55. In 
Figure 8 (b), the average PSNR performance of each 
algorithm is shown under real noise conditions. The 
ZZFC algorithm had the best PSNR values under various 
standard deviation conditions, indicating its excellent 
performance in image denoising. ZZFC significantly 
improved compared to the average PSNR of FCM and 
ZFC algorithms at noise standard deviations of 35, 45, 
and 55. The PSNR results of different algorithms under 
Gaussian noise interference are shown in Figure 9. 
 
0
20
10
140
80
ZFCM
110
FCM
ZFC
Average SSIM
(a) Gaussian noise (b) True noise
0
20
10
140
80
ZFCM
110
FCM
ZFC
Average SSIM
 

Basketball Video Image Segmentation Using Neutrosophic Fuzzy… Informatica 48 (2024) 151–154 151 
Figure 9: Different PSNR under gaussian noise interference 
 
In Figure 9 (a), the SSIM values obtained by the 
ZZFC algorithm were better when the noise standard 
deviation was set to 80, 110, and 140. The improvement 
rates of ZZFC reached 7.12%, 12.83%, and 18.29%, 
respectively, compared with the average SSIM of the 
FCM algorithm at noise standard deviations of 80, 110, 
and 140. The improvement of ZZFC was more significant, 
with 43.23%, 45.19%, and 41.18%, respectively, 
compared with the average SSIM of the ZFC algorithm 
under the same noise standard deviation. In Figure 9 (b), 
the SSIM values of the ZZFC algorithm were all optimal 
when the noise standard deviations were 80, 110, and 140. 
This meant that the denoised image had a higher 
structural similarity with the original image. The 
improvement percentages of ZZFC algorithm were 
11.23%, 18.27%, 21.09%, and 39.72%, 37.84%, and 
46.32%, respectively, compared with the average SSIM 
of FCM and ZFC algorithms at noise standard deviations 
of 80, 110, and 140. The ZZFC algorithm had the best 
denoising ability, followed by the FCM algorithm, which 
had a relatively weak denoising effect. The comparison of 
PSNR results for segmentation performance under 
different algorithms against salt and pepper and Gaussian 
noise is shown in Table 2. 
 
Table 2: Comparison of PSNR results for segmentation performance of different algorithms against spicy salt and 
gaussian noise 
Salt and pepper noise FCM ZFC ZZFC 
35% 11.15 13.96 14.96* 
45% 9.07 14.25 14.81* 
55% 7.79 13.88 14.57* 
Gaussian noise FCM ZFC ZZFC 
80 12.08 12.98 13.97* 
110 11.63 12.11 12.87* 
140 9.76 11.52 12.06* 
 
Note: * indicates a significant difference (P<0.05) 
between the ZZFC and FCM algorithms, while # 
indicates a significant difference (P<0.05) between the 
ZZFC and ZFC algorithms. 
According to Table 2, the ZZFC algorithm achieved 
the highest PSNR values at 35%, 45%, and 55% noise 
levels under salt and pepper noise, which were 14.96, 
14.81, and 14.57, respectively. There was a significant 
difference (P<0.05) between the ZZFC and FCM 
algorithms. Under Gaussian noise, the ZZFC algorithm 
also showed the best PSNR results at noise of 80, 110, 
and 140, with values of 13.97, 12.87, and 12.06,  
 
 
respectively. There was a significant difference (P<0.05) 
between the ZZFC and FCM algorithms. The ZZFC 
algorithm showed superiority in image segmentation 
performance in both environments. In Chapter 4, 
experiments were conducted on simple scene video 
segmentation using Tennis and Stefan video sequences to 
verify the algorithmic effectiveness. Tennis had relatively 
small background changes, while Stefan had relatively 
large background changes and noise. Figure 10 shows the 
average accuracy, recall, and F-value curves of the 
segmentation results. 
 
Accuracy
5 15 25 35
Frame rate
Split value
10 20 30
(a) Accuracy, Recall, and F-value Curve of 
Tennis Video Sequence Segmentation Results
0
0.2
0.4
0.6
0.8
1
5 15 25 35
Frame rate
Split value
10 20 30
(a) Accuracy, Recall, and F-value Curve of 
Stefan Video Sequence Segmentation Results
0
0.2
0.4
0.6
0.8
1
Recall
F-value
Accuracy
Recall
F-value
 
Figure 10: Average accuracy, recall, and F-value curves of segmentation results using different algorithms 
 
In Figure 10, the accuracy of Tennis video sequences was generally high, while the accuracy of 

152   Informatica 48 (2024) 152–154                                                                 X. Li
 
et al. 
Stefan video sequences slightly declined. In Figure 10 (a), 
the fluctuation of the segmentation curve was relatively 
small, with accuracy, F-measure, and recovery fluctuating 
between 0.8 and 0.97. In Figure 10 (b), Stefan's 
segmentation results decreased compared to the Tennis 
video sequence, but its accuracy only decreased by 4.6% 
in more complex scenes. The image background of the 
Tennis video sequence was relatively stable, and the 
overall effect of background removal was excellent. 
However, the video segmentation results of ZZFC 
algorithm showed higher recovery rate, accuracy, and 
F-measure compared with the segmentation results of 
Stefan video sequences. The comparison of accuracy, 
recall, and F-value results of the ZZFC algorithm under 
the Tennis and Stefan sequences is shown in Table 3. 
 
Table 3: Comparison of accuracy, recall, and F-value results of ZZFC algorithm under tennis and stefan sequences 
Video Sequence Algorithm Accuracy Recall F value 
Tennis 
ZFC 0.81 0.79 0.83 
ZFC 0.91 0.84 0.88 
ZZFC 0.96 0.92 0.93 
Stefan 
ZFC 0.37 0.29 0.93 
ZFC 0.23 0.26 0.92 
ZZFC 0.18 0.15 0.97 
 
Table 3 shows the comparison results of ZZFC, ZFC, 
and FCM algorithms in terms of accuracy, recall, and 
F-value. The ZZFC algorithm outperformed the other two 
algorithms in all aspects in the video sequence, with an 
accuracy of 0.96, a recall of 0.92, and an F-value of 0.93. 
Although the accuracy and recall of the ZZFC algorithm 
were 0.18 and 0.15, respectively, lower than ZFC and 
FCM algorithms for the Stefan sequence. The F-value 
was 0.97, still better than the other two algorithms. The 
ZZFC algorithm had higher performance in 
comprehensive evaluation. The parallel computing and 
preprocessing of video images could be considered to 
improve the computational efficiency for the 
shortcomings of the proposed algorithm. Otherwise, the 
algorithm parameters were set reasonably and adjusted 
according to different situations to improve the 
algorithmic performance. In addition, hardware 
acceleration such as FPGA can be utilized to significantly 
improve the algorithmic running speed and real-time 
processing capabilities. 
5 Discussion 
This study proposes a basketball video image 
segmentation study using the ZZFC algorithm. The 
experimental results showed that the average PSNR value 
of the ZZFC algorithm performed best when the noise 
standard deviation was set to 35, 45, and 55 in a salt and 
pepper noise environment. The average PSNR 
improvement ratio of ZZFC under the same noise 
standard deviation was 13.45%, 11.54%, and 9.19%, 
respectively, compared with the FCM algorithm. The 
average PSNR improvement ratio of ZZFC under the 
same noise standard deviation was 21.94%, 21.74%, and 
23.68%, respectively, compared with the ZFC algorithm. 
The average PSNR value of the ZZFC algorithm was still 
optimal under real noise conditions, proving its excellent 
performance in image denoising. When the noise 
standard deviation was set to 80, 110, and 140, the SSIM 
value of the ZZFC algorithm was better than other  
 
algorithms, indicating that the denoised image had a 
higher structural similarity with the original image. The 
image segmentation efficiency of our research algorithm 
was higher compared with references [6] and [10]. This is 
because our study combines the image segmentation 
method of neutral fuzzy mean clustering algorithm and 
neutral fuzzy theory, which can better handle the 
uncertainty of images. The C-means clustering algorithm 
was introduced to achieve more accurate image 
segmentation. The computational difficulty of the 
algorithm in this study was smaller compared with 
reference [11]. Similar to references [7] and [12], the 
algorithm in this study had poor image segmentation 
performance in complex backgrounds. In the future, 
further exploration can be made to combine it with other 
image processing techniques to improve the algorithm's 
image segmentation performance and efficiency. 
6 Conclusion 
Basketball video image segmentation is used in sports 
video analysis, event detection, and athlete tracking. 
However, the task of image segmentation becomes 
extremely difficult due to the complexity of basketball 
game scenes. These scenes involve multiple moving 
athletes, basketball, and environmental backgrounds. 
Traditional image segmentation algorithms often struggle 
to accurately segment athletes, basketball, and the 
background effectively, especially in noisy environments. 
Therefore, this study proposed a basketball video image 
segmentation study using the ZZFC algorithm to deal 
with this challenge. The proposed basketball video image 
segmentation method had important practical significance 
for improving image quality and visual effects. This study 
shows excellent performance in denoising and image 
similarity. However, the proposed algorithm still has poor 

Basketball Video Image Segmentation Using Neutrosophic Fuzzy… Informatica 48 (2024) 153–154 153 
efficiency and stability when dealing with large-scale 
data or complex scenes. These challenges may affect the 
application effect of the algorithm in actual basketball 
video image segmentation. Therefore, further exploration 
should be conducted to combine the ZZFC algorithm 
with other image processing techniques to improve the 
depth and breadth of image processing in future research. 
The monitoring system can more accurately detect and 
track various events and dynamics in basketball games by 
applying the proposed algorithm to monitoring and sports 
analysis. Therefore, accuracy and efficiency can be 
improved. The proposed algorithm can also be used to 
analyze and identify various actions and events in 
basketball games. This can help teams and coaches to 
have a deeper understanding of the game situation and 
player performance, thereby formulating more scientific 
tactical strategies. However, the proposed basketball 
video image segmentation method may involve 
infringement of privacy information. Therefore, 
fuzzification can be used when using algorithms to 
protect privacy. 
Reference 
[1]  F. Yu, Y. Zhu, X. Qin, Y. Xin, D. Yang, and T. 
Xu, “A multi-class COVID-19 segmentation 
network with pyramid attention and edge loss in 
CT images,” IET Image Processing, vol. 15, no. 
11, pp. 2604-2613, 2021. 
https://doi.org/10.1049/ipr2.12249 
[2]  W. Sprague, and E. R. Azar, “Integrating 
acceleration signal processing and image 
segmentation for condition assessment of asphalt 
roads,” Canadian Journal of Civil Engineering, vol. 
49, no. 6, pp. 1095-1107, 2022. 
https://doi.org/10.1139/cjce-2021-0116 
[3] H. Yin, N. Chen, L. Yang, and J. Wan, “Pop‐net: 
A self‐growth network for popping out the salient 
object in videos,” IET Computer Vision, vol. 15, 
no. 5, pp. 334-345, 2021. 
https://doi.org/10.1049/cvi2.12032 
[4]  H. Li, and J. Wang, “CAPKM++2.0: An upgraded 
version of the collaborative annealing power k 
-means++ clustering algorithm,” 
Knowledge-Based Systems, vol. 262, no. 28, pp. 
2-14, 2023. 
https://doi.org/10.1016/J.KNOSYS.2022.110241 
[5]  X. M. Long, Y. J. Chen, and J. Zhou, 
“Development of AR experiment on 
electric-thermal effect by open framework with 
simulation-based asset and user-defined input,” 
Artificial Intelligence and Applications, vol. 1, no. 
1, pp. 52-57, 2023. 
https://doi.org/10.47852/bonviewAIA2202359 
[6]  M. Fiaz, M. Z. Zaheer, A. Mahmood, L. Seung-Ik, 
and J. Soonki, “4G-VOS: video object 
segmentation using guided context embedding,” 
Knowledge-Based Systems, vol. 231, no. 14, pp. 
2-14, 2021. 
https://doi.org/10.1016/j.knosys.2021.107401 
[7]  Y. Cao, L. Sun, C. Han, and G. Jian, 
“Attention‐based video object segmentation 
algorithm,” IET Image Processing, vol. 15, no. 8, 
pp. 1668-1678, 2021. 
https://doi.org/10.1049/ipr2.12135 
[8]  Y. Zhang, W. Li, and Yang. P, “Surveillance 
video motion segmentation based on the 
progressive spatio‐temporal tunnel flow model,” 
Electronics Letters, vol. 57, no. 13, pp. 12-186, 
2021. https://doi.org/10.1049/ell2.12186 
[9]  R. Zhu, Y. Cui, E. Hou, and J. Huang, “Efficient 
detection and robust tracking of spermatozoa in 
microscopic video,” IET Image Processing, vol. 
15, no. 13, pp. 3200-3210, 2021. 
https://doi.org/10.1049/ipr2.12316 
[10]  Y. Wang, L. Wang, W. Zhang, F. Yang, and H. Lu, 
“Temporal consistent portrait video 
segmentation,” Pattern Recognition, vol. 120, no. 
2, pp. 108-143, 2021. 
https://doi.org/10.1016/j.patcog.2021.108143 
[11]  Z. Yang, P. Xu, Y. Yang, and B. Kang, “Noise 
robust intuitionistic fuzzy C-means clustering 
algorithm incorporating local information,” IET 
Image Processing, vol. 15, no. 3, pp. 805-817, 
2021. https://doi.org/10.1049/ipr2.12064 
[12]  Z. Dong, Y. Men, Z. Li, Z. Liu, and J. Ji, “Chilling 
injury segmentation of tomato leaves based on 
fluorescence images and improved k-means 
clustering,” Transactions of the ASABE, vol. 64, 
no. 1, pp. 13-22, 2021. 
https://doi.org/10.13031/trans.13212 
[13] K. G. Dhal, A. Das, S. Ray, and J. Galvez, 
“Randomly attracted rough firefly algorithm for 
histogram based fuzzy image clustering,” 
Knowledge-Based Systems, vol. 216, no. 15, pp. 
2-20, 2021. 
https://doi.org/10.1016/j.knosys.2021.106814 
[14]  E. Bas, and E. Egrioglu, “A fuzzy regression 
functions approach based on Gustafson-Kessel 
clustering algorithm,” Information Sciences, vol. 
592, no. 5, pp. 206-214, 2022. 
https://doi.org/10.1016/j.ins.2022.01.057 
[15]  K. Li, J. Xu, T. Zhao, and Z. Liu, “A fuzzy 
spectral clustering algorithm for hyperspectral 
image classification,” IET Image Processing, vol. 
15, no. 12, pp. 2810-2817, 2021. 
https://doi.org/10.1049/ipr2.12266 
[16]  Z. Zhao, S. Zhao, and J. Shen, “Real-Time and 
light-weighted unsupervised video object 
segmentation network,” Pattern Recognition, vol. 
120, no. 4, pp. 110-120, 2021. 
https://doi.org/10.1016/j.patcog.2021.108120 
[17]  Z. Zhao, Y. Jin, J. Chen, B. Lu, C-F. Ng, Y-H. Liu, 
Q. Dou, and P. Heng, “Anchor-guided online 
meta-adaptation for fast one-shot instrument 
segmentation from robotic surgical videos,” 

154   Informatica 48 (2024) 154–154                                                                 X. Li
 
et al. 
Medical Image Analysis, vol. 74, no. 2, pp. 
102-113, 2021. 
https://doi.org/10.1016/j.media.2021.102240 
[18]  J. Zhang, L. Zhu, L. Yao, X. Ding, D. Chen, H. 
Wu, Z. Lu, W. Zhou, L. Zhang, P. An, B. Xu, W. 
Tan, S. Hu, F. Cheng, and H. Yu, “Deep 
learning-based pancreas segmentation and station 
recognition system in EUS: development and 
validation of a useful training tool (with video) 
(vol 92, pg 874, 2020),” Gastrointestinal 
Endoscopy, vol. 93, no. 3, pp. 780-781, 2021. 
https://doi.org/10.1016/j.gie.2020.04.071 
[19] T. Zhang, D. Li, Y. Cai, and Y. Xu, 
“Super-resolution reconstruction of noisy video 
image based on sparse,” Informatica, vol. 43, no. 3, 
pp. 414-420, 2019. 
https://doi.org/10.31449/inf.v43i3.2916 
[20]  G. L. Franklin, E. P. Bacheladenski, D. C. B. 
Rodrigues, and A. C. S. Crippa, “Teaching video 
neuroImage: stop-motion chorea in PURA 
syndrome,” Neurology, vol. 100, no. 10, pp. 
492-493, 2023. 
https://doi.org/10.1212/WNL.0000000000201605 
[21]  F. Chen, H. Liu, Z. Zeng, X. Zhou, and X. Tan, 
“BES-Net: Boundary enhancing semantic context 
network for high-resolution image semantic 
segmentation,” Remote Sensing, vol. 14, no. 7, pp. 
1638-1638, 2022. 
https://doi.org/10.3390/rs14071638