Expert Systems With Applications 241 (2024) 122792

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

Heuristic-based image stitching algorithm with automation of parameters for

smart solutions✩
Katarzyna Prokop, Dawid Połap ∗
Faculty of Applied Mathematics, Silesian University of Technology, Kaszubska 23, Gliwice, 44-100, Poland

ARTICLE INFO ABSTRACT

Keywords: The analysis of two-dimensional images enables the detection of structures, objects and their further classi-
Image stitching fication. In the Internet of Things, images are data obtained from cameras or projections of signals. In this
Heuristic paper, we propose a hybridization of the classic RANSAC approach with a selected heuristic algorithm like
Internet of Things
Grey Wolf Optimizer. This solution allows for increasing the ability to adjust images by analyzing the detected
Drone application
key points of the image. Such hybridization is interesting in terms of the possibility of obtaining stitching
Image processing
RANSAC
accuracy. Moreover, we have introduced the automation of parameter selection to increase the accuracy and
Grey Wolf Optimizer speed of the real-time algorithm. For this purpose, an analysis of the results of the algorithm for combining
video frames into one large image allowed for a significant reduction in the amount of data processing time.
The proposed algorithm was tested on various heuristic algorithms, where their differences and impact on
the operation of the method were indicated and discussed. Moreover, the proposition was tested in practical
application by automatically creating an image based on video frames recorded by the drone.

1. Introduction points are called key points, which represent specific image features
and can be detected in different images.
Feature extraction is one of the most important tasks in terms of The most popular image-merging algorithms are SIFT (Pavan &
data processing (Anand, Gupta, Koundal, & Singh, 2023). In most cases, Jyothi, 2022) and ORB (Zhang et al., 2022). However, new algorithms
raw data contains a lot of unnecessary information or even noise (Ti- are also created as an alternative to known solutions. An example
wari et al., 2023). The consequence of leaving the data in this form is Nie, Lin, Liao, Liu, and Zhao (2022), where the authors proposed
means that artificial intelligence methods or even statistical techniques to use a convolutional neural network to image rectangling. Another
will process a lot of insignificant or distorted data. Extracting only the idea was based on the creation of a different matrix based on the
relevant information contributes not only to increasing the quality of given overlapping regions. Then, create the search region and find the
data, but also to minimizing the time of processing, or even learning optimal seam (Chen, Li, Peng, Wang, & Gong, 2022). Algorithms work
methods. Hence, feature extraction is a key element at the analysis on the principle of finding key points, their reduction and matching.
stage. Even in the case of image classification, such action is important After matching the points, the images are superimposed and the final
to minimize the amount of data or even to pre-process it. image is cropped. To this day, newer methods of operation are being
Image analysis is not only classification but also processing to built in order to obtain the best results in the shortest possible time.
achieve noise reduction, detection or even segmentation (Dhiman et al., In this paper, we focus on proposing an algorithm combining a heuris-
2022; Schapiro et al., 2022). Quite often a series of images is taken, tic approach based on selected algorithms and features extracted by
from which one is finally selected. Sometimes it also happens that RANSAC (Bolles & Fischler, 1981).
something will be moved, e.g. the camera. Then, as much data as pos- Our proposition is composed of integrating heuristic algorithms to
sible is extracted from the series of images. This is done by combining find the best matching of the characteristics of the images in order to
images into one in such a way that the final effect of the combination is stitch them. Our goal was to find a way to organize the process of
not visible. Unfortunately, such imposition of images is not the easiest searching for a homography between two images. This comes down to
task from the point of view of implementation. The reason is the way searching the keypoint space to find a limited number of points that will
serve as a ‘‘skeleton’’ to build an effective homography model. By an
the images are superimposed and the choice of specific places. These

✩ This work is supported by the Silesian University of Technology, Poland, No. BKM-661/RMS2/2022.
∗ Corresponding author.
E-mail addresses: [email protected] (K. Prokop), [email protected] (D. Połap).

https://doi.org/10.1016/j.eswa.2023.122792
Received 13 August 2023; Received in revised form 17 November 2023; Accepted 27 November 2023
Available online 1 December 2023
0957-4174/© 2023 Elsevier Ltd. All rights reserved.
K. Prokop and D. Połap Expert Systems With Applications 241 (2024) 122792

effective homography model, we mean a model that allows two photos possible solutions, an example of which can be, among others, the
to be stitched together to reflect reality as closely as possible. In other analysis of histograms (Liu et al., 2022). Zhou et al. (2022) analyze
words, our motivation was to reduce the randomness of the RANSAC the panoramic assessment of image quality. As part of the research,
algorithm while improving the obtained results. This is what the use they pay attention to obtaining the greatest possible visual impression,
of heuristic algorithms makes possible — searching the space further which is possible when the combined images have the smallest defects
in a very random way but still using knowledge about the best units in the form of distortions. For this purpose, they presented a method
of the heuristic algorithm that look for the optimal solution. Such an based on two modules: reference generation and quality prediction. The
improved algorithm can be widely used, especially in the field of IoT, conducted research indicates outstanding evaluation results, enabling
where each, even small, improvement affects the efficiency of devices. such methods in future technologies. Image stitching can be also made
Moreover, we propose a mechanism for the quick selection of the best by manifold optimization (Zhang & Huang, 2022). It is a solution
parameter settings to automate this process. The proposal is dedicated that uses commonly known transformations as matrix manifold. The
to smart solutions that use cameras to take pictures or videos. In the difficult environmental conditions in taking images can be quite prob-
classic approach, each image or frame is processed separately. Our lematic when combining data due to rapid changes in local features.
proposal is the analysis of samples and their combination in order to This topic was taken up in Jiang, Zhang, Fan, and Liu (2022), where
reduce the number of images being processed at a later stage. The main the authors suggested using an adaptive neural network to analyze
contributions of this paper are: features. Also, fuzzy logic is adapted to this task. It was shown in Lu,
Huo, and Cheng (2022), where the stitching method was composed by
• developing a new hybrid algorithm for estimate parameters of a
increasing contrast by analyzing histograms. Then, calculate matching
mathematical homography model,
points and reduce their number. The remaining points are used to
• combination of features of the RANSAC algorithm and a selected
calculate distance and gray differences in pixels — these are the inputs
heuristic algorithm to achieve improved image stitching results
in fuzzy inference. This allows for calculating the fusion weight used in
compared to the original RANSAC algorithm,
stitching. Another example is the processing image by edge detection
• a method of automating the selection of parameters in the algo-
algorithm and processes images by two branches neural network (Nie,
rithm based on the analysis of the results of the operation,
Lin, Liao, & Zhao, 2022).
• framework for minimizing the number of calculations in smart
The analysis of current solutions in the field of image stitching
devices by reducing the samples obtained from sensors recording
is significant in further processing. Such processing is found through
the image,
the use of artificial intelligence methods. An example is convolutional
• practical application of the proposed method in IoT solution using neural networks (CNNs) that learn to detect appropriate features by
a drone. finding various parameters of graphic filters. Unfortunately, it is not
The rest of the paper is composed of an analysis of existing knowl- always possible to prepare a database that will allow for the high
edge to outline current work in image processing. Next, the hybridiza- efficiency of such methods. An alternative to classical learning is the
tion proposal with the selected heuristic algorithm is described. This federated learning process. It is a way of training the network on small,
section also shows the possibility of automating the selection of param- private bases and sharing the network model. Having multiple networks
eters. The next section describes the possibility of using the proposed trained on different data, they are aggregated for better generalization
solution in the intelligent processing of graphic data. Section five of the network. Recent years have brought many different solutions in
describes the tests performed and presents a discussion of the results this area, which is significant for the modern Internet of Things. An
obtained. In the last section, we summarize the modeled system and example of the use of federated learning is medical facilities, where pa-
the results of the conducted research. tient data is sensitive, but can be used for local network training. Malik,
Naeem, Naqvi, and Loh (2023) proposed a framework for training X-
2. Related works ray image classification networks. Another use is autonomous vehicles,
an example of which was presented by Parekh et al. (2023). In the
Image analysis is one of the key directions of research due to the research paper, the security aspect was analyzed and improved by
potential and presence of images in almost every object. Let us take introducing gradient encryption. Zhang, Xue, Zhao, and Wang (2023)
the example of a smartphone that can use cameras on the front and used federated learning for fault diagnosis which was an interesting
back of the phone. The results are images and video files that can be application in the Internet of Things. The massive development and use
used in further analysis. In the era of the Internet of Things, objects of federated learning can be an excellent approach to the practical use
have various sensors that collect data for analysis. Hence, an important of deep learning methods that can be trained during implementation.
element is the use of various approaches to feature extraction and
their classification. An important aspect is to stitch the images into 3. Image stitching by RANSAC and heuristic
one if we have a series of images. An example is the use of a classic
approach, i.e. analyzing features with the SIFT algorithm (Wang, Liu, & In this section, a proposal to hybridize RANSAC with a selected rep-
Zhang, 2022). The authors proposed using this method by implement- resentative of heuristic algorithms is described. In order to better visu-
ing an alternative selection strategy to minimize the effect of geometric alize the operation of the proposed method, a simplified visualization,
distortion when combining several images. The SIFT algorithm can as well as pseudocodes for reimplementation, are shown.
be also improved by analyzing the features by Euclidean metric and
interpolation which was shown in Yan and Ma (2022). Another ap- 3.1. Random sample consensus
proach is to combine images using feature detection algorithms to
construct groups of triangles (Du et al., 2022). An interesting way of To solve the problem of fitting a model to the data given in com-
combining is the interpolation of homography and the transformation puter vision assignments, the Random Sample Consensus (RANSAC)
of similarity (Zhang & Huang, 2022). Studies show that it is possible to algorithm was created (Bolles & Fischler, 1981). The procedure being
increase the quality of the image connection by performing surgery on discussed carries out the division of the data set of inliers, elements that
the matrix. Again in Zhang et al. (2022), the classic ORB algorithm is can be calculated with the obtained model and outliers, which do not fit
modified by introducing a Gaussian pyramid. The presented modifica- this model. It can be also described as the process of repeated random
tion indicates a reduction in the number of calculations and introduces sub-sampling. Naturally, the aim is to estimate a model that produces
the variability of the scale which was one of the key problems of the the largest possible set of inliers. This provides for the possibility to use
original algorithm. The use of various algorithms is also one of the this approach in finding a homography between images.

2
K. Prokop and D. Połap Expert Systems With Applications 241 (2024) 122792

RANSAC algorithm is iterative, thus the number of iterations is inlier number maximization/outliers number minimization. In other
one of the hyperparameters required to complete the procedure. Other words, subsequent starting points have the possibility to move towards
values that determine the results being obtained are the minimal num- these elements that yielded the best (greatest) score so far.
ber of elements needed to define model parameters, a threshold that For the realization of this idea, the Grey Wolf Optimizer (GWO)
stands for accepting the data as well-fitted and the number of well-fitted algorithm was used. The main assumption of this optimization method
elements required to recognize the model as well-chosen. Additionally, is to treat every point in the considered space as a wolf belonging
the operation might be cut off if the result of set precision is achieved to the pack. This animal species lives according to the leadership
earlier, so the iteration number is maximal but can be reduced. hierarchy. To put it differently, four major levels of this hierarchy can
The operation of the algorithm starts by randomly selecting 𝑘 points be distinguished: starting with the most important group, they are 𝛼, 𝛽,
necessary to create a model. Then, a proposal of such model 𝐻 is 𝛿 and 𝜔 types of wolves. It is assumed, that the 𝜔 wolves are the elder
generated, the parameters of which are adjusted to the chosen points. and sick ones, so they do not contribute essential knowledge to the
With the obtained parameters, the remaining elements are checked if search for the considered problem solution. The reason for this situation
they fit the model along with the established threshold 𝑡. Subsequently, is the fact that determining the optimal result is based on the wolves’
the score 𝑠 of the model is determined to evaluate if it was well adapted hunting strategy. Because of that, the potential solution depends on the
to the given data. The score 𝑠 stands for the length of the inliers set. In behavior of the pack leader, the individual marked as 𝛼, as it leads the
situation when the score does not exceed the required number of fitted herd and is always the closest to the victim which is mentioned as the
elements 𝑑, the model is no longer considered. Otherwise, it is assumed wanted result. Also, the position of 𝛽 and 𝛿 wolves contribute to the
to be the best-created model so far. By later iterations, in which the operation of the method.
For the calculations, 𝛼, 𝛽 and 𝛿 are understood as single individuals.
maximum number of them is 𝑛, the newest generated proposition is [ ]
compared with the current best model based on scores. In the end, Each wolf is described by an 𝑟-dimensional vector 𝑥 = 𝑥1 , … , 𝑥𝑟 ,
the model parameters which are characterized by the best data fitting depending on the size of the given computational problem. This can be
are returned. As mentioned earlier, there is also a possibility to end understood as the position of each wolf. The initial pack is generated
the operation sooner than achieving the maximum allowed number of randomly, where the domain can be defined. An inherent element of
iterations by applying the condition defining the expected quality of this method is the regular selection of the best three wolves from the
the created model. The basic algorithm assuming an iterative approach pack because their movement affects the movements of the remaining
population units. This choice is based on each unit fitness comparison.
with no additional stop condition is also described in Algorithm 1.
Thus, for the problem of some function maximizing, the best fitness
means that the values that make up an individual provide the highest
Algorithm 1: RANSAC algorithm. value for this function among all other options. In this case, the fitness
Input: set of keypoints, number of iterations 𝑛, number of function for an individual 𝑥 = [𝑥1 , … , 𝑥𝑟 ] can be expressed through a
points 𝑘 to define a model, threshold 𝑡, number of formula:
points 𝑑 that makes a model well-fitted
𝑓 (𝑥) = |{𝑘𝑝 ∶ ‖𝑘𝑝 − 𝑘𝑝𝐻 ‖2𝐹 < 𝑡}|, (1)
Output: best model 𝐻𝑏𝑒𝑠𝑡
1 𝑖 ∶= 0; where 𝑘𝑝 means keypoint, 𝑘𝑝𝐻 is a keypoint gained by use of model
2 while 𝑖 < 𝑛 do 𝐻 generated for individual 𝑥, ‖ ⋅ ‖𝐹 stands for Frobenius norm and is
3 Select 𝑘 sample points randomly from the set; a measure of error between a real key point and a calculated one with
4 Fit a model 𝐻 to the selected points; model 𝐻.
5 Calculate 𝑠 score, the number of points from the set that Accordingly, after generating the pack, the three best-matched in-
fits 𝐻 (inliers) with threshold 𝑡; dividuals are selected, and information about their position enables
6 if 𝑖 == 0 then further calculations.
7 𝐻𝑏𝑒𝑠𝑡 :=𝐻; GWO is also an iterative algorithm. Assuming 𝑚 stands for the
8 𝑠𝑏𝑒𝑠𝑡 :=𝑠 number of iterations to perform, in every 𝑖th (𝑖 = 0, … , 𝑚 − 1) time
9 if 𝑠 > 𝑑 then step some additional coefficients have to be calculated: 𝑎𝑖 , 𝐴𝑖𝑙 and 𝐶𝑙𝑖 ,
10 if 𝑠 > 𝑠𝑏𝑒𝑠𝑡 then 𝑙 = 1, 2, 3. The equations for all these coefficients are as follows:
𝐻𝑏𝑒𝑠𝑡 ∶= 𝐻; 𝑎 − 𝑎𝑚𝑖𝑛
11
𝑎𝑖 = 𝑎𝑚𝑎𝑥 − 𝑚𝑎𝑥 ⋅ 𝑖, (2)
12 𝑠𝑏𝑒𝑠𝑡 ∶= 𝑠; 𝑚

13 𝑖 + +; 𝐴𝑖𝑙 = 2 ⋅ 𝑎𝑖 ⋅ 𝑝, (3)
14 return 𝐻𝑏𝑒𝑠𝑡 ;
𝐶𝑙𝑖 = 2 ⋅ 𝑝, (4)

3.2. Hybridization with heuristic algorithm where 𝑎𝑚𝑎𝑥 and 𝑎𝑚𝑖𝑛 stand for a range of argument 𝑎 which usually
decreases from 2 to 0 during the algorithm operation and 𝑝 is a random
value from the range [0, 1] which is redrawn on each subsequent occur-
RANSAC algorithm has been enhanced with a heuristic to improve
rence. The obtained values are then used in the consecutive components
obtained results in feature selection (see Fig. 1). This extended method
of the movement equation:
is used to calculate more efficiently a common model for key points de-
termined with a key point detector such SIFT algorithm (Lowe, 1999). 𝐷𝛼𝑖 = |𝐶1𝑖 ⋅ 𝑥𝑖𝛼 − 𝑥𝑖 |, (5)
The most important assumption is to use the knowledge about the
currently best-found global solution. Examples of algorithms that meet 𝐷𝛽𝑖 = |𝐶2𝑖 ⋅ 𝑥𝑖𝛽 − 𝑥𝑖 |, (6)
this premise are Particle Swarm Optimization (Kennedy & Eberhart,
1995) and Grey Wolf Optimizer (Mirjalili, Mirjalili, & Lewis, 2014).
𝐷𝛿𝑖 = |𝐶3𝑖 ⋅ 𝑥𝑖𝛿 − 𝑥𝑖 |, (7)
The movement of individuals in these algorithms is dependent on the
leaders’ localization. RANSAC enables to finding of the best model by
iterative approach by checking the quality of models suited to randomly 𝑋1𝑖 = 𝑥𝑖𝛼 − 𝐴𝑖1 ⋅ 𝐷𝛼𝑖 , (8)
selected points. The modification with some heuristics aims to stream-
line searching by a reasonable selection of starting points through the 𝑋2𝑖 = 𝑥𝑖𝛽 − 𝐴𝑖2 ⋅ 𝐷𝛽𝑖 , (9)

3
K. Prokop and D. Połap Expert Systems With Applications 241 (2024) 122792

Fig. 1. Visualization of the proposed RANSAC hybridization with a heuristic algorithm.

Algorithm 2: Grey Wolf Optimizer.

𝑋3𝑖 = 𝑥𝑖𝛿 − 𝐴𝑖3 ⋅ 𝐷𝛿𝑖 , (10) Input: wolves number 𝑤, iterations number 𝑚, range of 𝑎
Output: individual 𝛼
assuming that 𝑥𝑖𝛼 , 𝑥𝑖𝛽 and 𝑥𝑖𝛿 represent the positions of current 𝛼, 𝛽 and 1 Generate initial pack with 𝑤 individuals;
𝛿 wolves and 𝑥𝑖 is a position of wolf under consideration. 2 𝑖 ∶= 0;
Finally, the formula for the movement is as follows: 3 while 𝑖 < 𝑚 do
𝑋1𝑖 + 𝑋2𝑖 + 𝑋3𝑖 4 Calculate 𝑎𝑖 , 𝐴𝑖𝑙 , 𝐶𝑙𝑖 coefficients according to the equations
𝑥𝑖+1 = . (11) (2), (3), (4);
3
5 Calculate the fitness of each individual;
Therefore, each iteration entails moving each of the wolves closer and
6 Find the best individuals 𝛼, 𝛽, 𝛿;
closer to the victim, i.e. the solution sought. The successive operations
7 for every wolf do
that make up this algorithm are presented in Algorithm 2.
8 Calculate distances 𝐷𝛼𝑖 , 𝐷𝛽𝑖 , 𝐷𝛿𝑖 from the best individuals
As it was mentioned, RANSAC improvement by heuristic algorithm
aims to select subsequent starting points in a more orderly manner, according to the equations (5), (6), (7);
depending on the results obtained. In this case, 𝑘-element set of 9 Calculate coefficients 𝑋1𝑖 , 𝑋2𝑖 , 𝑋3𝑖 according to the
points (from previously found key points) needed to define a model equations (8), (9), (10);
in RANSAC is treated as an individual of the heuristic algorithm, so a 10 Update the position according to the equation (11);
wolf in GWO. The modified algorithm starts by randomly generating an 11 𝑖 + +;
initial pack consisting of 𝑤 𝑘-dimensional wolves. Then, for the whole
12 Find the best individual 𝛼;
pack, the following actions take place in the fixed number of iterations
13 return 𝛼;
𝑚:

• calculating needed coefficients (𝑎, 𝐴, 𝐶),

• calculating the score 𝑠 of each wolf - value for every proposed set
of 𝑘 points after building a model 𝐻 for it and its evaluating,

4
K. Prokop and D. Połap Expert Systems With Applications 241 (2024) 122792

• choosing the best three individuals 𝛼, 𝛽 and 𝛿, so the best three The second stage assumes the use of a calculated number for oper-
sample subsets of keypoints, ations on original images. This technique was adopted assuming that
• relocating every wolf according to the leaders’ position, if the if some points are so important on less detailed images, they will
move gives a higher score, be important in particular with a higher resolution. Another problem
solved by automation is the appropriate selection of the number of
remembering, that the score of each individual is a measure of the
iterations. To avoid repeating the algorithm when the result does not
quality of the model considered for it. Obviously, the movement of
improve significantly, we propose to add a condition to check whether
the individual cannot be accidental in this case. The model obtained
the score has changed sufficiently: if the difference between the newly
for points that are not key will not be satisfactory. Because of that,
generated model score and the actual best one is lower than 1% for the
after applying the equation of individual movement, the nearest key
next 5 iterations, the algorithm should stop. The difference 𝑑𝑓 between
point is chosen. A distance between considering key points can be
the two results can be presented as a formula:
measured by known distance metrics, for example, Euclidean metric
𝑠 −𝑠
in an appropriate dimension. 𝑑𝑓 = 𝑛𝑒𝑤 ⋅ 100%, (12)
When all iterations have been carried out, the leader 𝛼 is re-elected 𝑆
and the model created for it is returned. These steps are also presented where 𝑠𝑛𝑒𝑤 is a score for a model under consideration in a given
in Algorithm 3. iteration, 𝑠 is the current highest score and 𝑆 stands for the number
of all matches. The remaining parameters to be determined are the
error tolerance threshold 𝑡 and the number of points 𝑘 to define a
Algorithm 3: RANSAC improvement with GWO.
model. According to the conducted experiments, we propose setting the
Input: set of keypoints, number of points 𝑘 to define a model, value of 𝑘 to 4. Moreover, the error tolerance threshold 𝑡 determines
wolves number 𝑤, iterations number 𝑚, threshold 𝑡 the accurateness of the obtained model because based on its value the
Output: best model 𝐻𝑏𝑒𝑠𝑡 points are evaluated as fitting well. We recommend 0.3 as a universal
1 Generate initial pack with 𝑤 𝑘-dimensional individuals (sample value for this parameter.
points randomly selected from the set); Thus, the proposal of parameters automation can be expressed by
2 𝑖 ∶= 0; the following list of steps:
3 while 𝑖 < 𝑚 do
4 Calculate 𝑎𝑖 , 𝐴𝑖𝑙 , 𝐶𝑙𝑖 coefficients according to the equations 1. calculate the number of matches 𝑆0 for resized photos,
(2), (3), (4); 2. set 𝑘 = 4, 𝑡 = 0.3,
5 if i==0 then 3. performs the proposed algorithm with 𝑆0 individuals,
6 for every wolf do 4. check the difference 𝑑𝑓 between scores using the formula (12),
7 Fit a model 𝐻; 5. if the difference 𝑑𝑓 is smaller than 1% for 5 consecutive itera-
8 Calculate 𝑠 score with threshold 𝑡; tions, stop the algorithm.

9 Find the best individuals 𝛼, 𝛽, 𝛿 with the highest score;

4. Framework for the analysis of images in intelligent devices
10 for every wolf do
11 Calculate distances 𝐷𝛼𝑖 , 𝐷𝛽𝑖 , 𝐷𝛿𝑖 from the best individuals
The proposed algorithm may be part of other multi-functional
according to the equations (5), (6), (7); frameworks, where, in combination with other methods, can perform
12 Calculate coefficients 𝑋1𝑖 , 𝑋2𝑖 , 𝑋3𝑖 according to the specific image processing tasks, especially in the Internet of Things
equations (8), (9), (10); applications. Taking this into account, we present in this part a specific
13 Compute a movement according to the equation (11); example of the presented idea application.
14 Find the closest individual 𝑤𝑜𝑙𝑓𝑛𝑒𝑤 to obtained Technical solutions in the Internet of Things are based on one of two
movement; forms of data processing received by sensors. The first is the automatic
15 Fit a model 𝐻𝑛𝑒𝑤 to the 𝑤𝑜𝑙𝑓𝑛𝑒𝑤 ; processing of each image on a given device. However, there are also
16 Calculate 𝑠𝑛𝑒𝑤 score for 𝐻𝑛𝑒𝑤 model with threshold 𝑡 ; solutions based on sending data for analysis elsewhere, such as a cloud
17 if 𝑠𝑛𝑒𝑤 > 𝑠 then or server. Apart from the place of data processing (transmission time
18 Move 𝑤𝑜𝑙𝑓 to the 𝑤𝑜𝑙𝑓𝑛𝑒𝑤 position; and data security/privacy are the most important factors here), all
19 𝑖 + +; data is most often processed due to the potential benefits of analysis
and classification. In the case of processing data from one area that
20 Find the best individual 𝛼;
is moving, it is worth minimizing the number of calculations. An
21 𝐻𝑏𝑒𝑠𝑡 ∶= model fitted to 𝛼;
example would be a camera on a ship or in a car. Hence, the method
22 return 𝐻𝑏𝑒𝑠𝑡 ;
of processing will be an important element. We suggest using your
computing power to divide the video file into frames and subject them
3.3. Automation of parameter selection to automatic stitching using the previously proposed algorithm.
In order to minimize calculations, the algorithm will stitch frames
Due to the characteristics of both algorithms used, there are several 1 s apart. If stitching is not possible, the time between stitched frames
parameters whose values have a direct impact on the effectiveness will be reduced by half. When stitching turns out to be sufficient, fur-
of the method. For this reason, in this section, we propose a way to ther frames are similarly stitched. Of course, after combining a certain
automatically adjust them to obtain the most effective results. number of frames, the image is sent for further analysis, e.g. using
For automation of the method’s parameters, we propose a two-stage convolutional neural networks. A visualization of this process is shown
approach. The first step involves reducing the sizes of images but only in Fig. 2
for further calculations, thus the original images will be stitched. This
approach assumes that images are reduced to 10% of their original 5. Experiments
dimensions. Then, using the SIFT algorithm, key points are detected
and finally, the matches are determined with Lowe’s ratio test (Lowe, An inherent element of developing new solutions is subjecting them
2004). The obtained number of matches 𝑆0 will allow us to fit the to appropriate experiments. This section is intended for this purpose.
appropriate number of individuals in the proposed method. It has been divided into two parts, in which we first discuss the

5
K. Prokop and D. Połap Expert Systems With Applications 241 (2024) 122792

Fig. 2. Overview of the image stitching process in the practical application. Video frames are reduced, stitched together, and analyzed for completeness.

Fig. 3. Comparison of average image stitching times using different heuristic algorithms.

effectiveness of the proposed method itself, and then check how it optimizer (PBO) (Połap & Woźniak, 2017), artificial rabbits optimizer
performs in practical application. (ARO) (Wang, Cao, Zhang, Mirjalili, & Zhao, 2022), and wild horse
The proposed solution was tested by analyzing the proposed algo- optimizer (WHO) (Naruei & Keynia, 2022). Using the selected parame-
rithm of point analysis and stitching using images of size 3000 × 4000 ters, the processing with stitching times of the images were compared.
pixels. The operation of hybridization with the use of various heuristic The average measurement results are shown in Fig. 3. As part of the
algorithms was analyzed. Attention was paid to the influence of heuris- conducted tests, three variants of the coefficients were analyzed, such
tic coefficients. The tests were performed on a computer composed of as 20 iterations with 20 and 100 individuals and 100 iterations with
32 GB RAM, Intel Core i9-10850K 3.60 GHz, and NVIDIA GeForce RTX 20 individuals. It was noticed that the average times for all algorithms
3060. are similar. Of course, there are significant differences, as exemplified
Next, the focus was on the analysis of image stitching using the by the PBO algorithm. The reason for the longer processing time is the
found key points compared to the classical approach. The next step of additional mechanism of population reduction and replenishment. This
the research was to analyze the use of the algorithm in a simulated is an element that did not appear in the other heuristics. The more
solution in the field of the Internet of Things. individuals, the more operations are performed, as it can be seen by
using 100 with only 20 iterations. Apart from the PBO algorithm, the
5.1. Proposal analysis other algorithms had similar results with a large number of individuals
with a small number of iterations. However, GWO turned out to be
Analysis tests were undertaken on the basis of a set of pairs of 50 the fastest algorithm when it comes to stitching images. For only 20
images, where each pair was processed by the original algorithm and iterations and 20 individuals, GWO turned out to be a slower algorithm
the proposed solution. However, at the initial stage of the research, the than WHO. Hence, at small parameter values, WHO is the fastest one.
average running time of various heuristic algorithms was checked. Four Then GWO and ARO. However, image stitching is most often performed
different algorithms were selected: described here GWO, polar bear with large image sizes, so these parameters will be significantly larger.

6
K. Prokop and D. Połap Expert Systems With Applications 241 (2024) 122792

Fig. 4. Running time depending on iterations and individuals for GWO.

Table 1
Mean number of inliers for sample photos with 2318 key points in 15 launches. The highest result in each row is bolded. The best result overall is underlined.
Population Iterations ARO GWO PBO WHO RFOA RANSAC All iters.
10 10 498.53 518.13 438.00 – 475.07 5 1 8 . 3 3 100
10 20 550.07 589.93 463.80 – 583.47 574.07 200
10 30 584.53 622.67 416.60 – 586.60 596.80 300
20 10 509.40 584.87 496.67 248.27 537.93 574.07 200
20 20 571.93 648.53 436.40 229.47 562.47 593.53 400
20 30 640.53 624.07 452.47 186.20 519.53 639.53 600
30 10 582.33 643.60 473.13 365.60 414.87 596.80 300
30 20 592.33 682.33 498.47 304.73 420.60 639.53 600
30 30 632.60 676.80 428.93 282.73 374.27 659.40 900
40 10 554.27 647.33 480.40 301.67 258.13 593.53 400
40 20 642.67 657.40 472.46 324.40 257.13 651.33 800
40 30 664.33 677.20 471.00 310.67 307.20 681.47 1200

Heuristic algorithms differ from each other in ways of searching the tested using several recent algorithms, such as ARO (Wang, Cao et al.,
solution space. This is due to the different equations for the displace- 2022), GWO (Mirjalili et al., 2014), PBO (Połap & Woźniak, 2017),
ment of individuals in the population. However, it can be seen that the WHO (Naruei & Keynia, 2022), RFOA (Połap & Woźniak, 2021). The
overall operation is similar. The main idea of the heuristic algorithm object of the tests was a pair of very detailed photos, for which as many
guarantees a perfect solution provided that all possible solutions are as 2318 pairs of key points were detected using the SIFT algorithm.
checked. In the case of limiting the search by the number of individuals This allowed for the creation of comparative tables, as the difference
and iterations, it contributes to finding the optimal results. Hence, these with this number of key points was noticeable. It turns out that in the
algorithms will always return an optimal solution under given compu- case of more detailed images, the heuristic approach allows for a better
tational conditions, which will be more accurate with the increase of fit of the homography model than in the case of a completely random
these parameters. Hence, the choice of heuristics is important because RANSAC algorithm. Naturally, this approach is also associated with an
of the possibility of acting as quickly as possible and thus obtaining additional number of calculations performed in the course, hence the
optimal results in the shortest possible time.
proposed approach is slightly slower compared to the same number
In accordance with the presented results of time measurements for
of iterations. The first tally refers to the number of inliers that were
GWO, the focus was on the analysis of time in relation to individual
obtained. This is shown in Table 1. Each method was tested for a certain
heuristic parameters and how they affect the algorithm’s running time.
number of iterations and population size. In the case of the WHO
The results are shown in Fig. 4. These are the average values obtained
algorithm, due to its specificity, it was not possible to call methods for a
for 15 tests for each image from the test set. It is easy to see that
smaller number of individuals. For the most accurate comparison, in the
the measurement curves are almost linear, so the increase in the time
needed to process the image depends on the parameters. However, the case of the RANSAC algorithm, these two parameters were multiplied
average processing time at 10 iterations is slower to increase with more to obtain the total number of iterations (all iters.) that were performed
individuals. The difference in the average time for the application of in each approach. The table shows the average results for 15 calls of
40 individuals and 10 and 20 iterations is more than 5 s, which means each variant. 10 out of 12 comparisons performed favored the heuristic
that it is a two-fold increase in time. Assuming 40 individuals and 20 approach over the classical approach. In particular, it can be seen that
and 30 iterations, the difference was 3.5 s. Analyzing other cases, the the highest average is achieved for the GWO algorithm. The reason for
time difference decreases with increasing the number of iterations and this is probably the fact that this algorithm bases the search for space
keeping the same number of individuals. on information about the best individuals. This allows finding a model
During the experiments, a comprehensive comparison of the re- corresponding to more inliers than in the case of a completely random
sults achieved by the classical RANSAC algorithm and a heuristic- approach. In addition, the highest score of all variants also went to this
based proposition was conducted. The heuristic approach has been method. On the other hand, a comparison of the time consumption of

7
K. Prokop and D. Połap Expert Systems With Applications 241 (2024) 122792

Fig. 5. First example of comparison of RANSAC and Heuristic-based RANSAC.

Fig. 6. Second example of comparison of RANSAC and Heuristic-based RANSAC.

the tested methods was also prepared (see Table 2). Of course, for the effects of stitching samples for RANSAC and our proposal. The first
same number of iterations, the RANSAC algorithm is less demanding row shows the key point matches and the second row the stitching
because of no need to calculate unit movement equations. However, results. In Fig. 5, the number of found key points was higher in the
obtained differences are subtle, especially for GWO, WHO and ARO heuristic-based algorithm. In addition, there is much less shift in the
in comparison with RANSAC. Thus, further development of heuristic- image, and the moment of transition is much less noticeable. A similar
based approaches may result in even better results in a comparable situation occurs in the second case, which is shown in Fig. 6. Here it
time of operation. In particular, the right choice of heuristics for the should be noted that the main link element was a tree, which was more
specificity of the problem is important, as it can be seen in GWO results. thoroughly covered with key points for heuristics. However, a larger
The results of stitching sample images are shown in Figs. 5 and 6. area around it was covered by the original algorithm. In addition to the
For such large images, the number of iterations, which amounted to tree, other image elements have been selected as connecting objects.
200, turned out to be very important. In the images, we can see the For the original algorithm, these were much further house objects than

8
K. Prokop and D. Połap Expert Systems With Applications 241 (2024) 122792

Fig. 7. Close-ups of framed fragments of photos comparing the performance of the RANSAC algorithm and our approach.

Table 2
Mean duration of approaches for sample photos with 2318 keypoints in 15 launches. The shortest time in each row is bolded.
Population Iterations ARO GWO PBO WHO RFOA RANSAC All iters.
10 10 0.93 1.01 1.94 – 2.08 0.88 100
10 20 1.79 1.86 3.62 – 3.63 1.66 200
10 30 2.69 2.85 5.52 – 5.35 2.56 300
20 10 1.90 1.94 3.71 1.81 3.70 1.66 200
20 20 3.60 3.77 12.50 3.55 7.18 3.52 400
20 30 5.25 5.33 19.18 5.15 10.67 5.26 600
30 10 2.81 2.81 5.48 2.73 5.56 2.56 300
30 20 5.36 5.25 10.71 5.29 10.76 5.26 600
30 30 7.92 7.77 15.52 7.82 16.05 7.74 900
40 10 3.73 3.68 7.03 3.69 7.40 3.52 400
40 20 7.13 7.00 13.79 7.08 14.37 6.87 800
40 30 10.57 10.35 20.66 10.47 21.55 10.29 1200

for the second algorithm. It is worth noting that in Fig. 5, the area of when RANSAC is used. The heuristic-based approach here provides
points for the proposed algorithm was mainly on one object, although a better representation of the actual appearance of these objects. A
it was much more covered in area than when using the original one. similar phenomenon can be seen in the second set of photos. Using the
As a consequence, the hybridized algorithm allowed for more accurate proposed method provided a better way to connect the railing in the
localization of the features of the selected objects, which resulted in a foreground of the photo (see Fig. 7).
better match. In addition, the alignment of the images allowed for the
minimization of distortion caused by stitching. 5.2. Practical application of the proposal
Moreover, the fragments where the differences between the original
approach and our hybrid approach can be seen, have been marked by In order to analyze the proposed method in IoT, we have built a
a blue frame on the stitched photos. Fig. 𝑥 presents close-ups of the system that transmits a recorded file from a drone flight and finally
selected fragments. It can be seen that in the first test pair of photos, the creates an image composed of video frames. This solution is interesting
wall behind the bench is broken when RANSAC is used. Our approach due to the automatic ability to create an image of the entire surface over
resulted in the two parts of the wall being connected in a more gentle which the drone is moving. Such an image makes it possible to use it in
and fluid way. Additionally, the bench itself exhibits a similar problem the analysis of changes in the environment or data for creating maps.

9
K. Prokop and D. Połap Expert Systems With Applications 241 (2024) 122792

Fig. 8. Drone DJI Mavic 2 Pro used to analyze the selected area.

Fig. 9. An example of combining data collected using a drone.

In the conducted experiment, we used the DJI Mavic 2 Pro drone in Fig. 9. The presented image indicates a very good connection of
(see Fig. 8) which recorded the image while flying over the fields in the frames, which allows the generation of an image presenting the
Wodzisław Slaski in Poland. The recorded image was saved on the con- specified area. The figure also includes two indicated clippings where
nected Samsung S20 mobile device. After the recording was completed, the upper key points were shown using stitched five frames with the
the video file was uploaded to the server via an intermediary cloud. sixth one. A large number of points enabled very thorough punching of
The server automatically cuts the video file into frames. Then it only images, limiting distortions. However, minimal overlapping of images
left 1 frame per second. The images were stitched together using the (yellow line) was noticed, which is caused by applying the selected
proposed algorithm with GWO. The effect of the operation is shown frame. Despite this one element, the other photos are stitched without

10
K. Prokop and D. Połap Expert Systems With Applications 241 (2024) 122792

any connection signs or distortions. It is worth noting that the selected interesting and show another approach to the possibilities of processing
area presents low hills, rocks and dumps with similar features. The an image with stitching. Taking into account the achievements so far in
hybrid approach proposed in the article was used to combine frames improving this algorithm, our goal was to propose a method that would
from the recorded video. The heuristic part used 100 individuals and 20 control the unlimited randomness of the original algorithm and would
iterations — larger numbers may increase the accuracy of the results. work regardless of the nature of the input data.
However, automatic selection enabled the generation of an area whose
stitched frames are almost devoid of distortion, hence the process
6. Conclusions
constitutes a practical use of the proposed methodology for real-time
measurements in a smart environment.
In this paper, we proposed a hybridization of a classic algorithm
5.3. Comparison with existing methods like RANSAC with a selected heuristic representative. An example is
the selected Grey Wolf Optimizer, which made it possible to improve
There are a number of other algorithms in the literature whose aim the matching of points in two images, which contributed to more
was to improve the operation of the RANSAC algorithm and, ultimately, accurate stitching of the images. Performed tests on various heuristic
image stitching. For this reason, in this section, we decided to discuss algorithms showed that the selection of the heuristic algorithm depends
some of them that have appeared in recent years. Moreover, we would on parameters such as the number of iterations and individuals. The
like to point out their advantages, disadvantages and limitations. The undoubted advantage of the proposed solution is the ability to obtain a
works that present an improved RANSAC algorithm are mainly based visibly better point base for the searched homography in a comparable
on the appropriate elimination and selection of points for homography time. Although the original RANSAC algorithm performs slightly better
estimation. One emerging paper that follows this principle is Liu and Bu in the time comparison, the difference is actually small and the benefits
(2019). The presented approach is based on the elimination of outlier of using the hybrid method are clearly worth the cost. The marginally
points based on the Euclidean distance. In other words, only those pairs slower operation of the proposed algorithm is also completely natural,
of points that are neither too close nor too far apart are left for further as additional calculations result from the connection with the second
consideration, leaving only 60% of the points, instead of RANSAC algorithm. In turn, the limitations of this method result from the
randomly sampling all detected pairs. This approach allows narrowing characteristics of the algorithms used. RANSAC is completely random
down the number of calculations performed, but it is not a universal by design, while heuristic algorithms are also characterized by a certain
solution, because in special cases points indicating important features degree of randomness, but they also depend on the results ensuing from
may be omitted. For this reason, this solution provides faster, but less subsequent iterations of the search for a solution. Thanks to this, the
accurate operation. Another published work in this field is Canaz Sev- results obtained can only be better than the original ones. In addition,
gen and Karsli (2020), in which the authors randomly select points for the proposed mechanism can be effectively implemented in the Internet
which they find neighboring points and calculate the planimetric and of Things, where it can contribute to reducing the number of processed
height distances between a given one and its neighbor. Then, if a given images and enable obtaining more accurate results. Experiments carried
distance is within the adopted threshold, the neighboring point is con- out on a selected dataset of images have shown that such a solution
sidered a new cluster and its own neighbors are considered. However, significantly affects the results of time measurements. Moreover, the
it is worth adding that this solution has its specific application shown proposition can be easily used in the practical application of IoT
in this work: extracting points outside the plane boundary but within solutions. In this paper, we show that this proposal can be used for
the plane space for the problem of building roof planes. Therefore, it
creating an image composed of many frames during a drone’s flight for
is difficult to transfer this idea to a universal method for combining
the purpose of environmental analysis or map creation.
photos that may be of a completely different nature, although it serves
In future works, we want to focus on new methods of locating
its purpose in a given issue. Again, in Rahman, Li, and Yin (2019),
features and attempts to adapt the heuristic algorithm to minimize the
the idea was to reduce the randomness by introducing descending
number of these points. This solution may allow the analysis of more
likelihood-based minimal set selection. The mentioned solution is still
local and global features.
not a multipurpose technique, as it outperforms the RANSAC algorithm
only in the case of a highly noisy data set, as the authors point out in
experiments. CRediT authorship contribution statement
On the other side, when considering the latest achievements in the
field of photo stitching, solutions based on fuzzy logic and neural net- Katarzyna Prokop: Conceptualization, Methodology, Software, In-
works cannot be forgotten. Authors of paper (Shen, Ji, & Miao, 2019) vestigation, Writing. Dawid Połap: Conceptualization, Methodology,
created a model for real-time image stitching using Convolutional Writing.
Neural Networks. They point out that traditional methods of searching
for homography are very time-consuming, which cannot be criticized
Declaration of competing interest
for solutions based on a trained neural network. However, it cannot be
said that the neural network is a perfect solution in this case. Firstly,
the neural network must be properly prepared in advance, which also The authors declare that they have no known competing finan-
involves a lot of time. Each specific application of such a network also cial interests or personal relationships that could have appeared to
requires an appropriate training set, which should additionally contain influence the work reported in this paper.
many sample instances. In the case of stitching images, there are few
such collections, and most of them are adapted to specific sceneries. Data availability
Therefore, this type of solution also has a limited application. As shown
in the work (Lu et al., 2022), fuzzy logic can also be used to improve
No data was used for the research described in the article.
the results of photo stitching. The concept of this work is to reduce the
number of unmatched points by combining the orientations of feature
points. Then, the distance weight and gray differences of pixels in Acknowledgments
overlapping images are calculated. Both values are used as the input
data of fuzzy inference. The result of fuzzy control is an appropriate This work is supported by the grant at the Silesian University of
weight for adequate image fusion. The results obtained in this work are Technology, Poland No. BKM-505/RMS2/2023.

11
K. Prokop and D. Połap Expert Systems With Applications 241 (2024) 122792

References Nie, L., Lin, C., Liao, K., & Zhao, Y. (2022). Learning edge-preserved image stitching
from multi-scale deep homography. Neurocomputing, 491, 533–543.
Anand, V., Gupta, S., Koundal, D., & Singh, K. (2023). Fusion of U-net and CNN model Parekh, R., Patel, N., Gupta, R., Jadav, N. K., Tanwar, S., Alharbi, A., et al.
for segmentation and classification of skin lesion from dermoscopy images. Expert (2023). GeFL: Gradient encryption-aided privacy preserved federated learning for
Systems with Applications, 213, Article 119230. autonomous vehicles. IEEE Access, 11, 1825–1839.
Bolles, R. C., & Fischler, M. A. (1981). A RANSAC-based approach to model fitting and Pavan, M., & Jyothi, K. (2022). Human action recognition in still images using SIFT
its application to finding cylinders in range data. In IJCAI. Vol. 1981 (pp. 637–643). key points. In Intelligent systems and sustainable computing (pp. 313–325). Springer.
Canaz Sevgen, S., & Karsli, F. (2020). An improved RANSAC algorithm for extracting Połap, D., & Woźniak, M. (2017). Polar bear optimization algorithm: Meta-heuristic
roof planes from airborne lidar data. Photogrammetic Record, 35(169), 40–57. with fast population movement and dynamic birth and death mechanism. Symmetry,
Chen, J., Li, Z., Peng, C., Wang, Y., & Gong, W. (2022). UAV image stitching based 9(10), 203.
on optimal seam and half-projective warp. Remote Sensing, 14(5), 1068. Połap, D., & Woźniak, M. (2021). Red fox optimization algorithm. Expert Systems with
Dhiman, G., Juneja, S., Viriyasitavat, W., Mohafez, H., Hadizadeh, M., Islam, M. Applications, 166, Article 114107.
A., et al. (2022). A novel machine-learning-based hybrid CNN model for tumor Rahman, M., Li, X., & Yin, X. (2019). DL-RANSAC: An improved RANSAC with modified
identification in medical image processing. Sustainability, 14(3), 1447. sampling strategy based on the likelihood. In 2019 IEEE 4th international conference
Du, P., Ning, J., Cui, J., Huang, S., Wang, X., & Wang, J. (2022). Geometric on image, vision and computing (pp. 463–468). IEEE.
structure preserving warp for natural image stitching. In Proceedings of the IEEE/CVF Schapiro, D., Sokolov, A., Yapp, C., Chen, Y.-A., Muhlich, J. L., Hess, J., et al. (2022).
conference on computer vision and pattern recognition (pp. 3688–3696). MCMICRO: A scalable, modular image-processing pipeline for multiplexed tissue
Jiang, Z., Zhang, Z., Fan, X., & Liu, R. (2022). Towards all weather and unobstructed imaging. Nature Methods, 19(3), 311–315.
multi-spectral image stitching: Algorithm and benchmark. In Proceedings of the 30th Shen, C., Ji, X., & Miao, C. (2019). Real-time image stitching with convolutional neural
ACM international conference on multimedia (pp. 3783–3791). networks. In 2019 IEEE international conference on real-time computing and robotics
Kennedy, J., & Eberhart, R. (1995). Particle swarm optimization. In Proceedings of (pp. 192–197). IEEE.
ICNN’95-international conference on neural networks. Vol. 4 (pp. 1942–1948). IEEE. Tiwari, S., Jain, A., Sapra, V., Koundal, D., Alenezi, F., Polat, K., et al. (2023). A smart
Liu, J., & Bu, F. (2019). Improved RANSAC features image-matching method based on decision support system to diagnose arrhythymia using ensembled ConvNet and
SURF. The Journal of Engineering, 2019(23), 9118–9122. ConvNet-LSTM model. Expert Systems with Applications, 213, Article 118933.
Liu, Y., Wang, X., Sun, L., Chen, J., Lei, P., He, J., et al. (2022). Underwater image Wang, L., Cao, Q., Zhang, Z., Mirjalili, S., & Zhao, W. (2022). Artificial rabbits
stitching method based on optimal seam and multi-resolution fusion algorithms. In optimization: A new bio-inspired meta-heuristic algorithm for solving engineering
Eighth symposium on novel photoelectronic detection technology and applications. Vol. optimization problems. Engineering Applications of Artificial Intelligence, 114, Article
12169 (pp. 2759–2769). SPIE. 105082.
Lowe, D. G. (1999). Object recognition from local scale-invariant features. In Pro- Wang, S., Liu, C., & Zhang, Y. (2022). Fully convolution network architecture for
ceedings of the seventh IEEE international conference on computer vision. Vol. 2 (pp. steel-beam crack detection in fast-stitching images. Mechanical Systems and Signal
1150–1157). Ieee. Processing, 165, Article 108377.
Lowe, D. G. (2004). Distinctive image features from scale-invariant keypoints. Yan, S., & Ma, M. (2022). Image stitching based on an improved SIFT algorithm. In
International Journal of Computer Vision, 60, 91–110. ICETIS 2022; 7th international conference on electronic technology and information
Lu, J., Huo, G., & Cheng, J. (2022). Research on image stitching method based on science (pp. 1–5). VDE.
fuzzy inference. Multimedia Tools and Applications, 81(17), 23991–24002. Zhang, L., & Huang, H. (2022). Image stitching with manifold optimization. IEEE
Malik, H., Naeem, A., Naqvi, R. A., & Loh, W.-K. (2023). DMFL_Net: A federated Transactions on Multimedia.
learning-based framework for the classification of COVID-19 from multiple chest Zhang, Z., Wang, L., Zheng, W., Yin, L., Hu, R., & Yang, B. (2022). Endoscope image
diseases using X-rays. Sensors, 23(2), 743. mosaic based on pyramid ORB. Biomedical Signal Processing and Control, 71, Article
Mirjalili, S., Mirjalili, S. M., & Lewis, A. (2014). Grey wolf optimizer. Advances in 103261.
Engineering Software, 69, 46–61. Zhang, Y., Xue, X., Zhao, X., & Wang, L. (2023). Federated learning for intelligent
Naruei, I., & Keynia, F. (2022). Wild horse optimizer: A new meta-heuristic algorithm fault diagnosis based on similarity collaboration. Measurement Science & Technology,
for solving engineering optimization problems. Engineering with Computers, 38(Suppl 34(4), Article 045103.
4), 3025–3056. Zhou, Y., Gong, W., Sun, Y., Li, L., Wu, J., & Gao, X. (2022). Pyramid feature
Nie, L., Lin, C., Liao, K., Liu, S., & Zhao, Y. (2022). Deep rectangling for image stitching: aggregation for hierarchical quality prediction of stitched panoramic images. IEEE
A learning baseline. In Proceedings of the IEEE/CVF conference on computer vision and Transactions on Multimedia.
pattern recognition (pp. 5740–5748).

12