https://doi.org/10.31449/inf.v48i9.5872                                                          Informatica 48 (2024) 65-74   65 
 
  
Improved OTSU Theory of Image Multi-Threshold Segmentation by 
Incorporating Ant Colony Algorithm 
 
 
Guozhen Sang
1.*
, Xiaoyan Wang
1†
, Jianyang Zhang
1 
1
 School of Computer Science and Technology, Weinan Normal University, Weinan, Shaanxi, China,714099 
2
 School of Computer Science and Technology, Nanyang Normal University, Nanyang, Henan, 473061, China 
E-mail: wangxiaoyan20211@163.com, zjy57913@126.com, sstgzy@126.com 
 
Keywords: image segmentation, OTSU, ant colony algorithm, image segmentation, multi-thresholding 
 
Recieved: March 7, 2024 
In today's information age, people have higher requirements for image processing and classification and 
are increasingly concerned about the topic of multi-threshold image segmentation. In recent years, with 
the continuous development of computer technology and its theoretical basis, the relevant research 
direction gradually towards a more accurate and efficient direction. It is an algorithm that have both 
global and localized characteristics and can deal with dynamic changes in uncertain environments. This 
study introduces the enhanced OTSU theory and analyzes the improved multi-threshold image 
segmentation which is based on the information entropy and greedy factor and other related algorithms 
to segment the image, and can achieve two goals under this model: one is to maximize the grey value; the 
other is the optimal path. The author first analyzes the aging routes of ants and other experimental data 
and finds that there is not much discrepancy. Secondly, he finds a new method to reduce the competition 
between ant colonies and also reduce the noise pollution level, thus improving the overall performance 
and convergence speed. 
Povzetek: Članek predstavi izboljšano OTSU teorijo za večpragovno segmentacijo slik, ki vključuje 
algoritem kolonije mravelj. Predlagana metoda poveča natančnost in hitrost segmentacije s kombinacijo 
optimizacije mravelj in OTSU analize histogramov, kar izboljša prepoznavanje večpragovnih slik v 
različnih aplikacijah. 
 
 
1  Introduction 
Otsu theory has certain advantages in image segmentation, 
but there are also some shortcomings: on the one hand, 
because the method introduces a small and unstable 
amount of information, not easy to deal with noise and 
other defects leading to its inability to be applied to 
real-life scenes; on the other hand, although the improved 
algorithm overcomes the traditional ant colony algorithm 
for the same problem when there are different thresholds 
to choose different effects, still based on a single 
individual and although the improved algorithm 
overcomes the different effects of the traditional ant 
colony algorithm when selecting different thresholds for 
the same problem, it is still a difficult problem that cannot 
be solved by target detection and multi-stage search based 
on individual ant features [1-3]. 
Multi-threshold segmentation is a new image processing 
technique based on mathematical theory and computer 
graphics, which is based on the rational selection and 
optimization of objective functions and constraints. 
Although much progress has been made in the application 
of multi-threshold segmentation in image classification, 
there are still some problems: the experimental algorithm 
does not take into account the boundary conditions such as 
the amount of information and grey level changes between 
the target and the background [5-6]. Since the background 
is a feature of different regions in real situations, it is 
impossible to determine a suitable size range for 
comparison; at the same time, the image segmentation 
results of the three vertices of the target contour edge, 
edge, and bottom map obtained after binarization may 
also produce certain differences, thus leading to 
deviations in the segmentation effect. 
In the 1990s, domestic research on the fusion threshold 
segmentation algorithm began. First, some foreign 
scholars proposed and conducted experiments to verify 
that the method can effectively improve image quality and 
enhance imaging accuracy [7]; second, based on improved 
Otsu theory combined with Otsu theory to establish a 
concept model based on "multiple thresholds", 
"group-type coefficients The second is to establish a 

66   Informatica 48 (2024) 65–74                                                                                   G. Sang et al. 
conceptual model based on "multiple thresholds" and 
"group-type coefficients" to realize the extraction of 
information about the distance change between the target 
pixel and the background based on the improved Otsu 
theory [8] proposed a combined thresholding method for 
image processing, which improves the search efficiency in 
the target region by removing the noise. To address the 
traditional artificial grey-scale histogram blindly 
extracting features from the problem to achieve a new 
method of bilateral or multi-way cluster analysis, and used 
genetic algorithms to solve dynamic stochastic process 
parameter combination weights to obtain the global 
optimal solution with local optimal control effect [9]. 
 
In the development of image segmentation techniques, the 
main focus is on the analysis of the objective function to 
obtain a complete, effective, and realistic model of the 
system, a method that can be a good solution to the 
problems faced [10]. However, this method does not take 
into account the complex phenomena such as local 
optimal solutions and noise pollution arising from the 
influence of image greyscaling and topology, which 
makes it difficult to be widely used and promoted. The ant 
colony technique is computationally heavy and redundant, 
and the thresholds found are not global, making the 
algorithm somewhat unstable [11]. Therefore, it is 
improved by introducing the particle swarm algorithm 
into the ant colony algorithm and improving the fitness 
formula to achieve a fast, efficient, and accurate search for 
the best threshold, and also to effectively avoid the 
improved algorithm falling into a local optimum. 
 
2  Related works 
Table 1 depicts the comparison of previous studies. 
 
Table 1: Summary table in related work 
Ref Objective Method Limitation Result 
[12] Addressing noise 
interference and 
over-segmentation in 
current techniques is 
crucial for improving the 
accuracy of identifying 
diseases in maize foliar 
pictures. 
The complete particle 
swarm optimization (PSO) 
technique has been 
enhanced with a 
revolutionary elite 
approach to improve 
segmentation accuracy and 
optimal segmentation 
threshold. 
This method 
demonstrated enhanced 
segmentation 
efficiency, but it may 
face challenges in 
situations with 
unpredictable illness 
patterns or intricate 
environmental factors 
affecting picture 
quality. 
Experiments 
showed improved 
segmentation 
results compared 
to current 
algorithms on 
maize foliar 
disease photos, 
indicating the 
convergence and 
stability of the 
suggested 
method. 
[13] The development of a 
multi-threshold image 
segmentation approach 
aims to enhance the 
quality of medical image 
processing for more 
accurate diagnosis 
judgments. 
Create a horizontal and 
vertical crossover search 
ensemble 
multi-strategy-driven 
shuffled frog leaping 
algorithm (HVSFLA). 
The effectiveness of 
HVSFLA is 
significantly influenced 
by the quality of the 
original population and 
the configuration of its 
parameters, particularly 
in complex 
segmentation jobs or 
noisy photos. 
The experimental 
data reveals that 
HVSFLA 
outperforms other 
algorithms in 
terms of 
resilience and 
segmentation 
accuracy, 
indicating its 
potential for 
medical image 
segmentation 
applications. 
[14] To accelerate 
convergence and prevent 
local optima, provide an 
enhanced particle swarm 
optimization 
algorithm-based OTSU 
multi-threshold picture 
Employ a new method to 
compute particle 
contribution degrees, 
utilize asynchronous 
monotone learning factors 
to balance local and global 
search, and apply chaotic 
The intricacy of the 
picture data and the 
choice of parameters 
can affect how 
effective the procedure 
is. 
The proposed 
algorithm 
outperforms 
existing methods 
in segmentation 
accuracy, 
efficiency, and 

Improved OTSU Theory of Image Multi-Threshold Segmentation…              Informatica 48 (2024) 65–74     67      
segmentation technique. optimization to account for 
population variability. 
calculation times 
by 30%, 
demonstrating 
superiority over 
classic 
meta-heuristic 
algorithms. 
[15] To improve 
segmentation accuracy 
and efficiency, presented 
a hybrid whale 
optimization-based 
two-dimensional Otsu 
multi-threshold image 
segmentation method. 
The Otsu single-threshold 
technique will be expanded 
to two dimensions for 
multi-threshold 
segmentation, and a hybrid 
whale optimization 
technique will be 
implemented for improved 
accuracy. 
The suggested method's 
effectiveness may vary 
based on image 
parameters and 
attributes, and in some 
instances, 
computational 
complexity may still 
pose a challenge. 
The suggested 
approach 
outperforms 
conventional 
methods in 
segmentation 
efficiency and 
quality, 
surpassing Otsu 
multi-threshold 
segmentation in 
one dimension, as 
shown by PSNR 
and SSIM 
assessments. 
[16]  Reduce computing 
complexity, sluggish 
convergence, and 
premature convergence 
in current approaches to 
increase the precision of 
multi-threshold picture 
segmentation. 
The Adaptive Perturbation 
of Oscillation and 
Mutation Operation 
(AFOA-APM) 
Aptenodytes Forsteri 
Optimization Algorithm is 
suggested. 
Despite the 
enhancements, 
AFOA-APM may still 
encounter challenges 
with intricate photos or 
situations with uneven 
illumination. 
AFOA-APM 
outperforms 
existing methods 
in quality, 
consistency, and 
accuracy, as 
demonstrated by 
CEC 2017 
benchmarks and 
picture 
segmentation 
tasks, effectively 
addressing 
multi-threshold 
segmentation 
challenges. 
[17] A meta-heuristic 
equilibrium technique is 
being developed to 
enhance research quality 
and analysis by 
determining optimal 
thresholds in grayscale 
pictures for improved 
image segmentation 
accuracy. 
A unique meta-heuristic 
equilibrium technique was 
used to determine ideal 
threshold values using 
images from the Berkeley 
Segmentation Dataset, 
comparing performance 
against seven well-known 
techniques. 
This approach, despite 
excelling in greater 
levels of threshold can 
result in decreased 
average deviations and 
CPU time for fitness 
values, peak 
signal-to-noise ratio, 
and structural 
resemblance score. 
The new 
algorithm 
demonstrated 
superior 
performance in 
various criteria, 
including 
structural 
similarity index, 
maximum 
absolute error, 
peak 
signal-to-noise 
ratio, and fitness 
values, enhancing 
picture 
segmentation. 
[18] The study compares the 
effectiveness of AGDE, 
an enhanced Differential 
Evolution algorithm, 
with more advanced and 
With Rényi's entropy and a 
non-local mean 2D 
histogram, use AGDE in 
MTIS. Examine using the 
TPC and BSDS500 
Insufficient 
examination of various 
datasets and scant 
discourse on 
computational 
The AGDE-based 
segmentation 
approach, despite 
its superior 
performance 

68   Informatica 48 (2024) 65–74                                                                                   G. Sang et al. 
conventional methods in 
multi-threshold image 
segmentation, 
particularly for medical 
images. 
datasets. complexity. compared to its 
competitors, has 
the potential to 
enhance the 
accuracy and 
reliability of 
medical picture 
segmentation. 
[19] The use of an improved 
Whale Optimization 
Algorithm (LCWOA) 
has been utilized to 
enhance the quality of 
skin cancer picture 
segmentation and aid in 
disease detection. 
The WOA to enhance 
optimization speed, 
promote exploration, and 
verify CEC2014 
functionality for skin 
cancer image 
segmentation. 
The study focused on 
the performance 
comparison of WOA 
variations, specifically 
skin cancer pictures, 
and did not investigate 
its applicability beyond 
threshold segmentation. 
LCWOA 
outperforms 
current WOA 
variations in 
optimizing skin 
cancer images, 
enhancing 
efficiency, 
accuracy, and 
velocity, thus 
improving patient 
care and disease 
identification. 
 
3  Methods and materials  
 
3.1. Ant colony optimization algorithm 
The algorithm Ant colony optimization algorithm (ACO) 
is a meta-heuristic. Ants take the shortest route possible to 
locate food and get back to their colony, just as in real life. 
Ants employ pheromone trails to retrace their routes. In 
ACO, pheromones and fake ants are created to find the 
shortest path across a graph. Since pheromone evaporates, 
the ants will go along the path with the highest 
concentration of pheromone. ACO first constructs an 
initial solution from a finite set of solution components, 
just as in previous algorithms. The ant then proceeds to 
move down the graph, where each vertex denotes a 
component of the solution. The ACO was first proposed to 
be applied to the traveler's problem, There are n cities, the 
city i can be represented by a pair of real numbers (xi,yi) 
and its coordinates, and city j is represented by (xj,yj), then 
the distance between city i and city j is can be found by the 
following equation. 
 
𝑑 𝑖𝑗
=
√
(𝑥 𝑖 − 𝑥 𝑗 )
2
+ (𝑦 𝑖 − 𝑦 𝑗 )
2
               (1) 
 
 Then at moment t, the transfer probability P of choosing 
path dij is 
 
𝑃 𝑖𝑗
𝑘 (𝑡 ) = {
𝜏 𝑔 𝛼 (𝑡 )𝜂 𝑗 𝛽 (𝑡 )
∑ (𝑡 )𝜂 𝑎 𝛽 (𝑡 )
𝛼 𝑠𝑡𝑎𝑡𝑖𝑜𝑛𝑑𝑠𝑡𝑎𝑡𝑖𝑜𝑛𝑑 0
    ,            𝑗 ∈
𝑎𝑙𝑙𝑜 𝑤 𝑒 𝑑 𝑘 𝑜𝑡 ℎ𝑒𝑟𝑤𝑖𝑠𝑒            (2) 
 
The enlightening function mentioned above is 
 
𝜂 𝑖𝑗
=
1
𝑑 𝑠                                     (3) 
 
To avoid a situation where the amount of pheromone 
gradually disappears and too much residual pheromone 
overwhelms the heuristic information, after a traversal 
visit to all cities, the algorithm has to do a global 
pheromone update for this traversal, which is given by the 
following equation. 
 
𝜏 𝑖𝑗
(𝑡 + 𝑛 ) = (1 − 𝜌 )𝜏 𝑖𝑗
(𝑡 ) + 𝛥 𝜏 𝑖𝑗
(𝑡 ,
𝑡 + 𝑛 )    (4) 
 
The expression for the pheromone increment is: 
 
𝛥 𝜏 𝑖𝑗
(𝑡 ,
𝑡 + 𝑛 ) = ∑ 𝛥 𝜏 𝑖𝑗
𝑘 (𝑡 ,
𝑡 + 𝑛 )
𝑚 𝑘 =1
         (5) 
 
Under the ant-perimeter system model, ant-density system 
model, and ant-volume system model, the solution 
procedure for the parameters in equation (5) is as follows 
(6) (7) (8), respectively. 
𝛥 𝜏 𝑖𝑗
(𝑡 ,
𝑡 + 𝑛 ) = {
𝑄 𝐿 𝑘 0,   𝑜𝑡 ℎ𝑒𝑟𝑤𝑖𝑠𝑒          （6 ） 
𝛥 𝜏 𝑖𝑗
(𝑡 ,
𝑡 + 𝑛 ) = {
Q
0,   𝑜𝑡 ℎ𝑒𝑟𝑤𝑖𝑠𝑒         （7 ） 
𝛥 𝜏 𝑖𝑗
(𝑡 ,
𝑡 + 𝑛 ) = {
𝑄 𝐷 𝑖𝑗
0,   𝑜𝑡 ℎ𝑒𝑟𝑤𝑖𝑠𝑒         （8 ） 
 
The difference between the two is the different 
information emphasized. The flowchart of the steps of the 

Improved OTSU Theory of Image Multi-Threshold Segmentation…              Informatica 48 (2024) 65–74     69      
ant colony algorithm on the TSP problem is shown in 
Figure 1. 
 
Figure 1: Flowchart of the ant colony algorithm on TSP 
 
3.2. Otsu algorithm 
The total number of pixels in an image is represented by n. 
It is obvious that n has the following formula (9), 
𝑛 = 𝑛 1
+ 𝑛 2
+ ⋯ + 𝑛 𝐿                     (9) 
 
For each ni, the probability of its occurrence in the total 
number n is fi, then 
 
𝑓 𝑖 =
𝑛 𝑖 𝑛                              （10） 
∑ 𝑓 𝑖 𝐿 −1
𝑖 =0
= 1                         （11） 
Using the grey value i to partition the image into two 
classes, foreground, and background, represented using G₀ 
G₁ respectively, that is, (12), then define the proportion of 
pixels in these two classes G₀ and G₁ to be ω₀ and ω1. 
𝐺 0
= { 0
,
1
,
2
,
⋯
,
𝑡 }, 𝐺 1
= { 𝑡 + 1
,
𝑡 + 2
,
⋯
,
𝐿 − 1} （12） 
𝜔 0
= ∑ 𝑓 𝑖 1
𝑖 =0
                （13 ） 
𝜔 1
= ∑ 𝑓 𝑖 𝐿 −1
𝑖 =𝑎𝑖 +1
              （14） 
 
Assuming that ω₀ is ω1, then there is (15) and the total 
mean of the image is represented using μT 
𝜔 1
= 1 − 𝜔 0
             （15 ） 
𝜇 0
= ∑
𝑖 𝑓 𝑖 𝜔 0
=
𝑇 𝜔 0
𝑖 𝑖 =0
           （16） 
𝜇 1
= ∑
𝑖 𝑓 𝑖 𝜔 1
=
𝜇 𝑟 −𝑇 𝜔 1
𝐿 −1
𝑖 =𝑟 +1
      （17） 
𝜇 𝑇 = ∑ 𝑖 𝑓 𝑖 = 𝜔 0
𝜇 0
+ 𝜔 1
𝜇 1
  ,      𝑇 = ∑ 𝑖 𝑓 𝑖 𝑟 𝑖 =0
𝐿 −1
𝑖 =0
                         
（18 ） 
 
 The variances are respectively : 
𝜎 0
2
= ∑
(𝑖 −𝜇 0
)
2
𝑓 𝑖 𝜔 0
𝑖 𝑖 =0
          （19 ） 
𝜎 1
2
= ∑
(𝑖 −𝜇 1
)
2
𝑓 𝑖 𝜔 1
𝐿 −1
𝑖 =𝑖 +1
        （20 ） 
𝜎 𝑏 2
= 𝜔 0
(𝜇 0
− 𝜇 𝑇 )
2
+ 𝜔 1
(𝜇 1
− 𝜇 𝑇 )
2
（21 ） 
𝜎 𝑤 2
= 𝜔 0
𝜎 0
2
+ 𝜔 1
𝜎 1
2
            （22） 
𝑡 𝑞𝑢𝑡 = 𝑀𝑎 𝑥 𝑏 { 𝜎 𝑏 2
(𝑡 )}       （23 ） 
𝑡 𝑞𝑒𝑡 = 𝑀𝑖 𝑛 0≤𝑡 ≤𝐿 −1
{ 𝜎 𝑤 2
(𝑡 )}       （24 ） 
𝜎 𝑏 2
= 𝜔 1
𝜔 2
(𝜇 0
− 𝜇 1
)
2
=
[𝑇 −𝜇 𝑇 𝜔 (𝑡 )]
2
𝜔 (𝑡 )[1−𝜔 (𝑡 )]
  （25 ） 
 
Assuming that an image has a grayscale of L, the total 
number of pixels of the image can be expressed as: 
 
𝑁 = ∑ 𝑛 𝑖 𝐿 −1
𝑖 =0
                  （26） 
𝑓 𝑖 =
𝑛 𝑖 𝑁                 （27） 
 
For the multi-threshold Otsu algorithm, assuming that the 
image is to be segmented into n classes, it is obvious that 
there can be a total of n-1 segmentation thresholds to 
segment the image into n different classes. Let us assume 
that these n thresholds are denoted as T1, T2,..., Tn-1,n 
classes are obtained by segmenting the image using n-1 
thresholds, and for these n classes, the probability of 
occurrence of the total pixels in each class is assumed to be 
F1, F2,... ,Fn-1,where. 
 
𝐹 𝑘 = ∑ 𝑓 𝑖 𝑇 𝑖 +1
−1
𝑖 =𝑇 𝑘               （28） 
𝜇 𝑘 =
1
𝐹 𝑘 ∑ 𝑖 𝑓 𝑖 𝑇 𝑘 +𝑖 −1
𝑖 =𝑇 𝑘             （29 ） 
𝜎 𝑘 2
= ∑ (𝑖 − 𝜇 𝑘 )
2
𝑓 𝑖 𝐹 𝑘 𝑇 𝑘 +1
−1
𝑖 =𝑇 𝑘      （30） 
 
The mean grey value of the image, the inter-class variance, 
and the intra-class variance of each class are 
𝜇 = ∑ 𝑖 𝑓 𝑖 𝐿 −1
𝑖 =0
                     （31 ） 
𝜎 𝑏 2
= ∑ 𝐹 𝑖 (𝜇 𝑖 − 𝜇 )
2
𝑛 −1
𝑖 =0
            （32 ） 
𝜎 𝑤 2
= ∑ 𝐹 𝑖 𝜎 𝑖 2
𝑛 −1
𝑖 =0
                  （33） 
 
Multi-threshold Otsu optimal threshold should follow the 
same principle, and the largest or smallest set of thresholds 
is the multi-threshold Otsu optimal threshold. 
 
4   Improved OTSU theory of image 
multi-threshold segmentation 
incorporating ant colony algorithm 

70   Informatica 48 (2024) 65–74                                                                                   G. Sang et al. 
 
4.1 Main principles of the improved 3D 
OTSU algorithm 
The three-dimensional OTSU method is first optimized 
using the ant colony algorithm. The resulting suboptimal 
solution is then optimized using the chaos method. The 
improved algorithm captures the global nature more and 
avoids local extremes. The chaos optimization algorithm 
reflects the computational process of the algorithm as a 
search process for chaotic trajectories. To obtain a sizable 
partitioning effect, the given function variables are 
linearly mapped in the interval [0, 1]. In this paper, a 
logistic mapping is used which is simple in principle. The 
algorithm is mapped into chaotic sequence variables for all 
but the best solution by the following equation. 
𝑥 𝑘 0
+1
= 𝜇 ∗ 𝑥 𝑘 0
(1 − 𝑥 𝑘 0
)             （34 ） 
  
𝑋 = (𝑥 1
,
𝑥 2
,
⋯
,
𝑥 𝐷 ), 𝑋 (𝑘 0
) = (𝑥 1
𝑘 0,
𝑥 2
𝑘 0,
. . .
,
𝑥 𝐷 𝑘 0
)                             
（35） 
 
The main steps of the chaos optimization method are as 
follows. 
(1) Initialization 
Let 
𝑘 0
= 0, 𝑋 (𝑘 0
) = 𝑋 (0), 𝑥 ∗
= 𝑋 (0), 𝑓 (𝑥 ∗
) = 𝑓 (𝑋 (0))                      
(36) 
 
Transform the optimization variables into chaotic 
variables h within 0 to 1 according to the following 
formula.  
 
ℎ𝑥 𝑗 𝑘 0
=
𝑥 𝑗 𝑘 0
−𝑥 𝑗 𝑚𝑖𝑛 𝑥 𝑗 𝑚𝑎𝑥 −𝑥 𝑗 𝑚𝑖𝑛                     (37) 
 
(2) Calculate the chaotic variables for the next iteration 
Transform h into the chaotic variables for the next iteration 
according to the following equation: 
 
ℎ𝑥 𝑗 𝑘 0+1
= 𝜇 ℎ𝑥 𝑗 𝑘 0
(1 − ℎ𝑥 𝑗 𝑘 0
)               (38) 
 
(3) Solving for the optimization variables 
Translate h into the optimization variables according to the 
following equation: 
 
𝑥 𝑗 𝑘 0+1
= 𝑥 𝑗 𝑚𝑖𝑛 + ℎ𝑥 𝑗 𝑘 0+1
(𝑥 𝑗 𝑚𝑎𝑥
− 𝑥 𝑗 𝑚𝑖𝑛 )     (39) 
 
(4) Performance comparison 
Compare the performance of the resulting new solution by 
optimizing the variables. 
 
𝑥 (𝑘 0+1)
= (𝑥 1
𝑘 0+1,
𝑥 2
𝑘 0+1,
. . .
,
𝑥 𝑗 𝑘 0+1,
. . .
,
𝑥 𝐷 𝑘 0+1
)                           
(40) 
 
(5) Solution update 
If the new solution is better, then we have (41), otherwise, 
x* remains unchanged. 
 
𝑥 ∗
= 𝑥 (𝑘 0+1)
                        (41) 
 
(6) Output results 
Output x* when the maximum limit of chaotic searches is 
reached, otherwise return to step (2). The flow chart is 
shown below in Figure 2. 
 
 
Figure 2: Flow chart of the improved 3D OTSU method 
algorithm 
 
The paper suggests a unique method for image 
multi-threshold segmentation by merging ACO with 
Otsu's method. The advantages of both approaches are 
combined in this integration: Otsu's approach uses 
histogram analysis to effectively choose the best 
thresholds, and ACO offers strong optimization 
possibilities. The suggested strategy improves 
convergence characteristics and segmentation accuracy by 
combining various techniques. It has potential uses in 
several domains where accurate segmentation is essential, 
such as medical imaging and remote sensing. It could 
prove to be a useful tool for image analysis problems 
involving multi-threshold segmentation with more testing 
and improvement. 
 
4.2 Simulation experiments and analysis of 
results 

Improved OTSU Theory of Image Multi-Threshold Segmentation…              Informatica 48 (2024) 65–74     71      
 
Dataset 1: For fingerprint image 1 and fingerprint image 2, 
the original fingerprint image is relatively clear, and the 
segmentation results obtained by both algorithms are also 
clearly visible. For fingerprint image 3, the image of the 
bottom of the stripe is slightly blurred, resulting in an 
average segmentation result shown in Figure 3. In general, 
however, the fingerprint pattern is segmented by both 
algorithms. 
 
Figure 3: Fingerprint image segmentation results 
 
Dataset 2: This paper [20] makes use of the colonoscopy 
tissue segment data set from the Bio-Medical Grand 
Challenge. Out of 93 complete slide pictures of whole 
slide imaging (WSI), 250 images with an average size of 
5000 × 5000 pixels are included in the data set. The study 
creates a binary mask, where 255 represent the malignant 
tumor and 0 represents the backdrop. 
One of the drawbacks of the traditional thresholding 
method is that it does not accommodate multiple 
thresholding of the image. Table 1 below compares the 
regional consistency and runtime results of the two 
segmentation algorithms. Table 1 shows that the 
difference in time is not significant, but the regional 
consistency is significantly higher. The simulation proves 
that the DP artificial neural network with a combination of 
integrated whole-amplitude and bilateral filtering can 
effectively improve the problem of distributing the 
number of ants and the total energy concentration in the 
target region (i.e. the threshold can be well expressed 
when the number of particles exceeds a certain value in the 
same dimension). The improved algorithm introduces a 
chaotic local search algorithm to make the obtained 
thresholds more accurate. Therefore, the improved 
algorithm proposed in this paper can be used as a practical 
method for segmenting fingerprint images.
 
Table 1: Results of fingerprint image runs 
Picture Particle swarm optimization by the 3 D 
OTSU method 
The algorithm in this paper 
 Regional 
consistency 
time (ms) Regional 
consistency 
time (ms) 
Picture1 0.942 605 
 
0.947 
 
600 
Picture2 0.986 611 
 
0.995 
 
594 
Picture3 0.963 463 0.980 
 
460 
4.3. Comparison result 
 
The key processes have been developed using the Python 
3.11.4 environment. The suggested optimization 
techniques were tested in a simulated environment using a 
Windows 11 laptop equipped with an Intel i5 11th Gen 
CPU and 32 GB of RAM. This study used OTSU-ACO 
approaches for image multi-threshold segmentation to 
analyze data. In OTSU-ACO identification tasks, 
numerous statistics are utilized to measure a model's 
prediction capabilities, including Accuracy, f-measure, 
Sensitivity, Specificity, and Computational time. Particle 
swarm optimization (PSO) [20], Darwinian particle swarm 
optimization (DPSO) [20], and fractional order Darwinian 
particle swarm optimization (FODPSO) [20], 
One-dimensional OTSU [21], Two-dimensional OTSU 
[21], 2D OTSU-GA [21] were compared with our 
proposed method. Table 2 represented the overall 
comparison of the proposed method. 
 
Table 2: Outcomes of Accuracy, Sensitivity, F-Measure, Specificity 
Methods Accuracy Sensitivity F-Measure Specificity 
PSO [20] 0.76 0.77 0.55 0.72 
DPSO [20] 0.86 0.86 0.66 0.85 
FODPSP [20] 0.89 0.86 0.68 0.87 
OTSU-ACO [Proposed] 0.96 0.94 0.95 0.94 

72   Informatica 48 (2024) 65–74                                                                                   G. Sang et al. 
 
 
Accuracy: The accuracy with which pixels are accurately 
classified into several intensity levels is referred to as 
picture multi-threshold segmentation accuracy. It gauges 
how well-split areas and ground truth accord. Accurately 
classifying pixels into intensity levels is measured by 
segmentation accuracy, which is important for many 
image processing applications such as object identification 
and medical imaging. Figure 4 shows the comparison of 
accuracy.  
 
 
Figure 4: Outcome of accuracy 
 
The suggested OTSU-ACO approach achieves a better 
accuracy of 0.96, outperforming other methods like PSO, 
DPSO, and FODPSP. 
 
Sensitivity: Sensitivity quantifies the percentage of real 
positives that a segmentation model accurately identifies. 
It is essential for determining accuracy. View a 
multi-threshold segmentation example. Using intensity 
thresholds as a guide, multi-threshold segmentation 
separates a picture into many areas. It's helpful for 
precisely extracting items with different grayscale values. 
Figure 5 shows the comparison of sensitivity. 
 
 
Figure 5: Outcome of sensitivity 
In comparison to the other methods under comparison, the 
suggested OTSU-ACO approach has the maximum 
sensitivity of 0.94, outperforming the values of 0.77, 0.86, 
and 0.86 for PSO, DPSO, and FODPSP, respectively. 
F Measure: The F-measure, which combines precision 
and recall, is a statistical indicator of test accuracy. To 
assess segmentation algorithms, it offers a single score that 
strikes a balance between these two measures. Figure 6 
shows the comparison of F-measure.  
 
 
Figure 6: Outcome of F-Measure 
 
With an F-Measure of 0.95, the suggested OTSU-ACO 
approach outperforms PSO, DPSO, and FODPSP, which 
had scores of 0.55, 0.66, and 0.68, respectively. 
 
Specificity: The capacity of a segmentation algorithm to 
precisely detect true negatives is referred to as specificity. 
Multi-threshold segmentation needs to discern various 
items in a picture according to their intensity levels. Figure 
7 shows the comparison of Specificity. 
 
 
 
Figure 7: Outcome of specificity 
 
With a specificity score of 0.94, the suggested 
OTSU-ACO approach outperforms the other methods that 
were tested, with scores of 0.72, 0.85, and 0.87 for PSO, 
DPSO, and FODPSP, respectively.  
 
Computation time: The amount of time needed for a 
computer or algorithm to do a task is referred to as 
computational time. By dividing an image according to 
many intensity thresholds, a technique known as 
multi-threshold segmentation makes it easier to separate 
and analyze objects. Table 3 and Figure 8 show the 
computation time outcome. 

https://doi.org/10.31449/inf.v48i9.5872                                                          Informatica 48 (2024) 65-74   73 
 
 
Table 3: Comparison of computational time 
Computational Time 
 
No of images 
1 2 3 
One-dimensional OTSU [21] 8.236 7.994 8.727 
Two-dimensional OTSU [21] 12.1534 12.967 12.071 
2D OTSU-GA [ 21] 10.241 11.692 11.734 
OTSU-ACO [Proposed] 6.235 6.348 7.259 
 
 
Figure 8: Outcome of computational time 
 
The one-dimensional OTSU, two-dimensional OTSU, and 
two-dimensional OTSU-GA image segmentation 
approaches have computing times ranging from around 
7.994 to 12.967 seconds. The effectiveness of the 
suggested OTSU-ACO approach is enhanced, with 
timings ranging from 6.235 to 7.259 seconds. 
5  Discussion  
Multi-threshold segmentation techniques are essential in 
many domains, from agricultural disease detection to 
medical imaging. These researches showcase LCWOA 
strategies [18] for addressing issues like computational 
complexity and noise interference, such as multi-strategy 
ensembles and improved optimization methods. 
Notwithstanding these drawbacks, they show encouraging 
improvements in efficiency and accuracy, which are 
essential for accurate segmentation and diagnosis. 
Existing techniques such as (PSO) [20], (DPSO) [20], and 
(FODPSO) [20] have the following drawbacks: they are 
inefficient in handling large optimization landscapes, 
sensitive to parameter choices, and converge to local 
optima. These restrictions may result in less-than-ideal 
segmentation outcomes and higher computing expenses. 
By combining ACO for global optimization with Otsu's 
approach, which automatically generates appropriate 
thresholds, the suggested OTSU-ACO technique 
overcomes these shortcomings. The approach improves 
the segmentation robustness and accuracy by utilizing 
Otsu's efficiency in threshold selection and ACO's 
capacity to investigate a variety of solutions. It also 
improves adaptation to complicated picture properties and 
lessens susceptibility to parameter adjustments, which 
leads to better segmentation performance. This technique 
produces accurate segmentation results in a 
multi-threshold segmentation situation by iteratively 
adjusting threshold values depending on the image 
intensity distribution. This effectively divides an image 
into discrete sections. 
 
6  Conclusion 
In this thesis, take the threshold selection in the 
experiment as the research content, combine the research 
results of domestic and foreign scholars on image 
multi-threshold segmentation algorithm, work from the 
improved Otsu theory in moving average filter peak 
detection and recognition, fusion and model, etc., by 
proposing a new threshold segmentation method to 
achieve image multi-smoothing, and extend it to the image. 
In this paper, the main focus is on the segmentation and 
study of images based on an improved algorithm. In this 
paper, Otsu theory is used as a fusion of ant colony 
algorithm and thresholding to obtain the objective function. 
Firstly, the principle and implementation process is 
introduced, and a new idea of "multimodal", is proposed to 
deal with the constraints and boundary conditions, the 
optimal solution for noise pollution, and can improve the 
convergence characteristics, finally verify the effect and 
advantages, and finally draw the conclusion that the 
improved algorithm has good performance. Even though 
the suggested approach appears to be enhancing image 
segmentation, more research is necessary to determine 
whether it can be applied in practical settings and scaled to 
bigger datasets. Future studies may also look into using 
deep learning methods for improved efficiency and 
accuracy in segmentation across a range of imaging 
applications. 
 
Acknowledgment 
(1) The Humanities and Social Sciences Research Projects 
of the Ministry of Education,  

74   Informatica 48 (2024) 65–74                                                                                   G. Sang et al. 
Research on university learning behavior model based on 
the fusion of spatio-temporal trajectory data and 
multi-source data (NO: 19C10481026). 
(2) Key research and development and promotion projects 
in Henan Province,  
Research and development of medical endoscope system 
based on narrow band imaging and intelligent image 
fusion (NO: 192102310457). 
 
 
References 
[1] Tian Xian Z, SunL, Tian Zhen Z. Image 
reconstruction based on ant    colony algorithm [J]. 
Computer Science, 2020, 47( Suppl 2):231-235. 
[2] Shi, C., Zeng, Y. Y., & Hou, S. (2021). Summary of 
the application of swarm intelligence algorithms in 
image segmentation. Computer Engineering and 
Applications, 57(8), 36-47.. 
[3] Li Q, Lu Y, Lu R. High-precision segmentation of 
image contours based on proven ant colony 
algorithm [J]. Journal of Heilongjiang Institute of 
Technology 
(ComprehensiveEdition ),2021,21(1):47-51. 
[4] Chen C. Boundary extraction of medical images 
based on improved ant colony algorithm [J]. 
Computer Applications and Software, 2019, 
36(10):227-232. 
[5] Cao M. Research on image segmentation system 
based on improved ant colony algorithm [J]. Journal 
of Xi'an College of Arts and Sciences (Natural 
Sciences Edition ),2019,22(2):49-52. 
[6] Wang X fang, Zou Q, Peng L, et al. An ant colony 
image enhancement algorithm incorporating fuzzy 
clustering [J]. Data Acquisition and Processing, 2020, 
35(3):506-515. 
[7] Coello, C. A. C., & Zacatenco, C. S. P. (2012). List 
of references on constraint-handling techniques used 
with evolutionary algorithms. Information 
Sciences, 191, 146-168.. 
[8] Cao Z. Research on image segmentation system 
based on improved ant colony algorithm [J]. 
Electronic Design Engineering, 2019, 
27(20):166-170. 
[9] Liu X, Wang R, Chen C. Image segmentation based 
on improved ant colony algorithm [J]. Gansu Science 
and Technology, 2020,36(3):17-23. 
[10] Xiao Y. Research on the application of intelligent 
image recognition based on ant colony algorithm [J]. 
Computer Knowledge and 
Technology,2019,15(31):212-214. 
[11] Mohammdian-Khoshnoud M, Soltanian AR, 
Dehghan A, Farhadian M. Optimization of fuzzy 
c-means (FCM) clustering in cytology image 
segmentation using the gray wolf algorithm. BMC 
Molecular and Cell Biology. 2022 Feb 15;23(1):9. 
[12] Chen C, Wang X, Heidari AA, Yu H, Chen H. 
Multi-threshold image segmentation of maize 
diseases based on elite comprehensive particle 
swarm optimization and otsu. Frontiers in plant 
science. 2021 Dec 13;12:789911. 
[13] Chen Y, Wang M, Heidari AA, Shi B, Hu Z, 
Zhang Q, Chen H, Mafarja M, Turabieh H. 
Multi-threshold image segmentation using a 
multi-strategy shuffled frog leaping algorithm. 
Expert Systems with Applications. 2022 May 
15;194:116511. 
[14] Zheng J, Gao Y, Zhang H, Lei Y, Zhang J. OTSU 
multi-threshold image segmentation based on 
improved particle swarm algorithm. Applied 
Sciences. 2022 Nov 13;12(22):11514. 
[15] Ning G. Two-dimensional Otsu multi-threshold 
image segmentation based on      hybrid whale 
optimization algorithm. Multimedia Tools and 
Applications. 2023 Apr;82(10):15007-26.  
[16] Zhang P, Yang J, Lou F, Wang J, Sun X. 
Aptenodytes Forsteri optimization algorithm 
based on adaptive perturbation of oscillation and 
mutation operation for image multi-threshold 
segmentation. Expert Systems with Applications. 
2023 Aug 15;224:120058. 
[17] Abdel-Basset M, Chang V, Mohamed R. A novel 
equilibrium optimization algorithm for 
multi-thresholding image segmentation problems. 
Neural Computing and Applications. 2021 
Sep;33:10685-718. 
[18] Chen J, Cai Z, Heidari AA, Chen H, He Q, 
Escorcia-Gutierrez J, Mansour RF. 
Multi-threshold image segmentation based on an 
improved differential evolution: a case study of 
thyroid papillary carcinoma. Biomedical Signal 
Processing and Control. 2023 Aug 1;85:104893. 
[19] Liu L, Kuang F, Li L, Xu S, Liang Y. An efficient 
multi-threshold image segmentation for skin 
cancer using boosting whale optimizer. 
Computers in Biology and Medicine. 2022 Dec 
1;151:106227. 
[20] Kanadath A, Jothi JA, Urolagin S. Multilevel 
colonoscopy histopathology image segmentation 
using particle swarm optimization techniques. 
SN Computer Science. 2023 Jun 7;4(5):427. 
[21] Peng Z, Wang L, Tong L, Zou H, Liu D, Zhang C. 
Multi-threshold image segmentation of 2D 
OTSU inland ships based on improved genetic 
algorithm. Plos one. 2023 Aug 
25;18(8):e0290750. 

Improved OTSU Theory of Image Multi-Threshold Segmentation…              Informatica 48 (2024) 65–74     75