SPECIAL SECTION ON GIGAPIXEL PANORAMIC VIDEO WITH VIRTUAL REALITY

Received August 8, 2020, accepted August 25, 2020, date of publication September 7, 2020, date of current version February 2, 2021.
Digital Object Identifier 10.1109/ACCESS.2020.3022011

Active Contour Image Segmentation Method for

Training Talents of Computer Graphics and
Image Processing Technology
XINGHUO YE1 AND QIANYI WANG 2
1 School of Information Engineering, Huzhou University, Huzhou 313000, China
2 AI Lab Artificial Intelligence Laboratory, Beijing Chengshi Wanglin Information Technology Company Ltd., Beijing 100001, China

Corresponding author: Qianyi Wang ([email protected])

This work was supported in part by the Scientific Research Fund of Zhejiang Provincial Education Department, under Grant
Y201533314 and Grant JY24095, and in part by the Research Fund of Huzhou University, under Grant 201441.

ABSTRACT Image segmentation is a key technology in the field of computer image processing. Among
them, segmentation methods based on active contour models have been developed rapidly in recent years
due to their effective processing of complex images such as medical images. These methods have achieved
significant results in medical, military, and industrial fields. Present research work mainly introduces the
training of computer graphics and image processing technology and the method of active contour image
segmentation. It focuses on the study of image segmentation methods and focuses on the segmentation
methods based on active contour models. Firstly, it summarizes two types of segmentation methods based
on edge and region and summarizes their advantages and disadvantages. Then, the segmentation method
based on the active contour model is studied, and several typical active contour models are comprehensively
compared. Finally, the local binary fitting model and the local Gaussian distribution fitting energy model are
improved and simulated. Furthermore, from the development of computer graphics and image processing
technology to analyze some methods and means of training this professional talent. The experimental
results of this article show that the active contour image segmentation algorithm can not only ensure the
image segmentation algorithm but also reduce the number of iterations and shorten the image segmentation
time. Compared with the CV, LBF, and LGIF models computational efficiency of Segmentation method is
increased by 9.2 times, 2.64 times, and 1.44 times, respectively.

INDEX TERMS Computer graphics and image processing technology, talent training, active contour model,
image segmentation method.

I. INTRODUCTION methods are increasingly difficult to meet the fragmentation

In recent decades, people have done a lot of in-depth research requirements in practical applications such as medical imag-
on image segmentation technology, and hundreds of image ing and imaging.
segmentation methods have been widely studied and applied At present, image segmentation has a wide range of appli-
in various fields of computer vision. In fact, researchers cations in the fields of medical imaging, target recognition,
proposed image segmentation methods and corresponding industrial communities, and traffic control. In the field of
solutions for specific image research. So far, there is no uni- image processing, the results of image fragmentation can be
versal image segmentation method. Some traditional image used to observe the pathological changes of organs. At the
segmentation methods only use the information of the image same time, image segmentation plays an important role in
itself and strictly follow the bottom-up hierarchical cal- three-dimensional reconstruction, image recording, and med-
culation process. However, these traditional fragmentation ical diagnosis. It is used for the automation of remote sens-
ing images and synthetic bulkhead radar. Export the area
The associate editor coordinating the review of this manuscript and of interest. It is widely used in biometric recognition, such
approving it for publication was Zhihan Lv. as fingerprint recognition, IRIS recognition, and plays an

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 9, 2021 19187
 X. Ye, Q. Wang: Active Contour Image Segmentation Method for Training Talents

important role in traffic monitoring and vehicle target moni- Among them, F(C) is the curvature function of the evolu-
toring. Therefore, it is of great practical significance to study tion curve C, F(C) = ak, a is a constant, and k is the curvature
the image segmentation technology in more detail. of the curve C. This function is used to drive the evolution
Khadilkar et al. introduced the realization of a new auto- curve to the target boundary [5]. N represents the outer normal
matic face recognition system. Establishing an automated vector, and C(v, 0) = C0 (v) is the initial evolution curve.
framework for the recognition of human faces is a very Next, the level set method is applied to the above-
challenging task. It is also of great significance in related mentioned curve evolution equation expression. The
engineering disciplines such as computer graphics, pattern level set method embeds the evolution curve into the
recognition, psychology, image processing, and artificial neu- three-dimensional surface b(x,y,t), where b(x,y,t) is generally
ral networks. In response to this problem, authors also pro- It is expressed as a level set function with a signed distance
posed a side face authentication method based on discrete function, usually taken as:
wavelet transform and artificial neural network. Discussed
a subset determination strategy that expands the number of C(t) = {(x, y)|b(x, y, t) = 0} (3)
training samples and allows protection of global informa- That is, the zero level set function, the evolution equation
tion. However, the proposed face recognition system was of the geometric active contour model (GACM) is obtained
immature, making the recognition result not so accurate [1]. as follows:
Wu et al. proposed light field imaging technology that can
∂b ∇b
capture richer visual information from our world. Contrary = g(x, y)|∇b|[V0 + div( )] (4)
to traditional photography technology, the light field collects ∂t |∇b|
light radiation from all directions, thereby demultiplexing the Among them, V0 is the constant value speed, g(x,y) is the
angular information lost in traditional photography. However, gradient function of the image edge, the value of g(x,y) can
due to the immaturity of the technology, the data display faces be used to judge whether the pixel I(x,y) is on the edge of the
huge challenges [2]. Li G developed a fully automatic algo- image when the magnitude of the image gradient is relative
rithm based on machine learning to extract cracks from con- When it is smaller, the equation of g(x, y) ≈ 1, g(x,y) is as
crete bridge images. The algorithm combines an improved follows:
active image contour segmentation model based on regions 1
and a linear support vector machine using a greedy search g(x, y) = (5)
1 + |∇Gσ ∗ I (x, y)|2
strategy to eliminate noise. However, due to the inconsistent
strength of the cracks, the results will be biased [3]. Among them, Gσ represents σ a Russian filter with vari-
The innovation of this paper is to propose an improved ance Bbb, so the evolution process of the contour curve C is
matching algorithm. Aiming at the problem of feature the change process of the level set function b(x, y, t) [6].
description, the SICA algorithm is used to reduce the dimen- Because the level set method is used in the geometric active
sion of the feature section, which reduces the number and contour model, the geometric active contour model has more
dimension of feature distribution and reduces the amount advantages than the parameterized model. This method can
of calculation in the matching process. In response to the effectively deal with the topological structure changes of the
problem of high irregularity, directional restrictions have contour curve and solve the instability problem of numerical
been added under restricted conditions. The direction of the calculation. However, the disadvantage of this method is that
feature vector and the Euclidean distance is matched twice to it is prone to edge leakage. Since this method did not conduct
reduce the irregularity rate. It is used to find the end of the a deeper study on the energy function optimization of ACM
function during matching to ensure the time consumption of and other issues, it only combined the level set method and
the algorithm. curve evolution simply. On this basis, the researchers quickly
proposed many related improvement models [7].
II. ACTIVE CONTOUR IMAGE SEGMENTATION METHOD
A. ACTIVE CONTOUR MODEL ALGORITHM B. IMAGE SEGMENTATION METHOD
The geometric active contour model (GACM) is proposed on The edge is defined by the pixels in the image. When-
the basis of the combination of the level set method and curve ever the image intensity changes suddenly, an edge occurs,
evolution theory [4]. Let the image I (x, y) be the evolution and the corresponding pixel is defined as an edge pixel.
curve: Reflected in the objects in nature, that is, changes in the
boundaries, backgrounds, and materials of the objects lead to
C(v, t) = [x(s, t)y(s, t)] (1)
changes in brightness in the image [8]. Edge detection is the
where s is the normalized arc length of the curve, s ∈ [0, 1] most basic tool for image segmentation, and it plays a great
and t is time, the equation expression of the curve evolution role in the discovery, recognition, and segmentation of an
with time is: object in an image. This type of method detects the edges of
 ∂C(v, t)

segmented objects by finding points with similar properties.
= F(C)N
∂t (2) Edge-based segmentation methods are constantly being
C(v, 0) = C (v)
0 explored and improved. Among them, the segmentation

19188 VOLUME 9, 2021

X. Ye, Q. Wang: Active Contour Image Segmentation Method for Training Talents

methods based on edge detection are divided into two cat- among them:
egories, including search-based methods, namely the first
1 − x 2 +y22
derivative method, and zero-cross point methods, namely G(x, y, σ ) = − e 2πσ (12)
the second derivative method. The so-called search-based 2π σ 2
method mainly finds the boundary by deriving the first deriva- This is the two-dimensional Gaussian function used in
tive calculated from the image. The so-called zero-crossing smoothing [10].
method detects the zero-crossing through the second deriva- Step 2, then use the Laplacian operator to further manipu-
tive to measure the edge intensity [9]. Related methods are late the obtained smooth image:
introduced as follows:
∇ 2 g(x, y) = ∇ 2 (G(x, y, σ )∗ f (x, y)) = ∇ 2 G(x, y, σ )∗ f (x, y)
1) THE FIRST DERIVATIVE (13)
The gradient of this type of operator corresponds to the first
derivative. The gradient of a continuous function f (x, y) at The above calculation process is the Marr operator, also
the position (x, y) is expressed by the following formula: known as the LoG operator, which combines the Gaussian
filter and the Laplacian filter and combines the advantages of
∂f ∂f T the two. So, the edge detection effect of this operator is better.
∇f (x, y) = [Gx Gy ]T = [ ] (6)
∂x ∂y 3) Canny operator: It is an algorithm that uses a hierar-
Its amplitude and direction angle is: chical optimization algorithm to detect edges and is widely
used [11]. Similarly, the operator uses a specific Gaussian
1
mag(∇f ) = ||∇f(2) || = [G2x + G2y ] 2 (7) filter with a specified standard deviation to obtain the convo-
φ(x, y) = arctan(Gy /Gx ) (8) lution with a given image, smooth the image, and effectively
overcome the noise. The specific algorithm steps are:
The Prewitt operator mask is a discrete differential oper- Step 1: Use a specific Gaussian filter to process the original
ator. In actual calculations, two 3 × 3 masks are used to image to reduce the influence of noise, and its size should be
approximate the horizontal and vertical derivative values to fit flexibly changed with the image situation.
the true gradient value. This operator is more prone to noise. Step 2: Use a suitable first-order operator to detect the
The specific formula is as follows: gradient and direction of the filtered image.
    Step 3: In this step detailed processing is performed on the
−1 0 1 1 1 1
obtained edge information. The non-maximum elimination
Hx =  −1 0 1  Hy =  0 0 0  (9)
method is used to compare the center pixel with the neighbor-
−1 0 1 −1 −1 −1
ing pixels. When the central pixel is less than or equal to the
2) SECOND DERIVATIVE
neighborhood pixel value in the gradient direction, the central
pixel is set to zero [12].
1) Laplace operator: There must be a point when the zero-
Step 4: select dual thresholds to determine the final edge
crossing occurs, and this point can be used for edge detec-
detection result. The dual thresholds are two thresholds, one
tion. However, the outstanding feature of the second-order
large and one small. The pixels larger than the larger threshold
derivative operator is that it is relatively sensitive to noise.
are defined as strong edge pixels, and the pixels in the middle
Generally, in practical applications, only the function of the
of the two thresholds are defined as weak edge pixels. This
operator to determine the brightness and darkness of the pix-
step can further avoid the influence of noise, false edges, and
els is used, and it is not directly used for the initial detection,
loss of effective edges.
so as to avoid the double pixel width affecting the accuracy
of the result. The Laplace value of any point (x, y) is defined
C. IMAGE PROCESSING PRIORITY ALGORITHM
as follows:
Independent Component Analysis (ICA) is a data processing
∂ 2f ∂ 2f method. The basic principle is that there is a relationship
∇ 2f = + 2 (10)
∂x 2 ∂y X = AS for the observed entity S is an independent entity,
2) Marr operator: Compared with the Laplacian operator, and A is a complete classification table. Suppose there is a
this operator overcomes its shortcomings of being sensitive grade of Y, then Y = WX = WAS[13].
to noise. By using the zero-crossing point of the second In the expression, W is a mixing matrix. Similarly, WA is
derivative to realize the detection and determination of the an identification table, and the estimated vector (Y) is an
edge position, a wider range of applications. The specific independent vector. According to the central boundary the-
process of the algorithm is as follows: orem, the results show that the Gaussianness of the inde-
Step 1, filter the original image through a two-dimensional pendent body is stronger than the observed body. When the
Gaussian function, and obtain convolution to achieve smooth- non-Gaussian is strongest, the rating entity is closer to an
ing of the image and deal with noise: independent entity. The ICA algorithm uses the hybrid matrix
corresponding to the highest non-standard feature of X as the
g(x, y) = G(x, y, σ )∗ f (x, y) (11) projection matrix to achieve dimensionality reduction [14].

VOLUME 9, 2021 19189

 X. Ye, Q. Wang: Active Contour Image Segmentation Method for Training Talents

The result of processing with the PCA algorithm is very When processing Micro-CT medical slice images, the size
low. The PCA algorithm is largely affected by the number of the image is equal to the size of the clustering template,
of samples, and the characteristics of the output result in and then according to the template defined by the clustering,
the actual application are very unstable. The ICA algorithm the area where the target position is located in the image
assumes that the data are independent of each other and pro- is 1, and the area where the background is located in 0,
cesses the data by deriving high-level statistical data. At the which is initially obtained. The initial curve of the level set
same time, the sample size has no significant impact on SICA, of the image based on the clustering information. The initial
and the results of export characteristics are better. In terms curve of the level set is different from the previous rules.
of constraints, the PCA algorithm requires that each compo- The K-LGIF model not only effectively enhances the seg-
nent is uncorrelated, while the ICA algorithm requires strict mentation of the low contrast part of the Micro-CT medical
independence. From the data point of view, it is irrelevant, image, but also improves the recognition efficiency of the
but it is not necessarily independent, because independence segmentation of the image target.
should not be relevant. The ICA algorithm seems to reveal
the essence of the data more comprehensively. Structurally, B. GRAPHIC IMAGE RETRIEVAL EXPERIMENT
it makes more effective use of the basic features of the The present study designs an experiment to compare the
image [15]. advancement of retrieval methods based on the improved
This paper proposes an improved ICA algorithm. The PAM algorithm, and chooses the Coral image library as
idea of subtyping is added to the ICA algorithm, and the the image library to be retrieved in the comparison. There
SICA algorithm is proposed. The algorithm first processes are 20 main categories, each with 100 image targets, and a
the image into blocks to form a sub-model, and then uses the total of 2000 image targets are used for retrieval in experi-
ICA algorithm to process the sub-model to obtain a mixed ments [18]. The particle population is initialized in the PSO-
matrix, and finally selects the non-mixed matrix used by the FWA-PAM algorithm, the number of particles is set to 40,
image. The SICA optimization algorithm combines the idea w decreases linearly from 0.7 to 0.2, c1 = c2 = 0.4, r1 and r2
of sub-modes to minimize the interference of each target are random numbers between 0 and 1, the group updates the
sub-block change on the overall change of the target so that algebra N is set to 4, and the update algebra is set to 50.
even if the target has a significant interference impact, it can In this experiment, the PSO-K-Means image retrieval
only be reflected in the sub-block, and it will affect the target. method, K-means image retrieval method, and the method in
The overall feature extraction cannot have a major inter- this paper are used for image retrieval operations on the same
vention, which improves the algorithm’s robustness to local computer, and the performance of each algorithm is evaluated
changes. The main steps of the SICA optimization algorithm by the accuracy of each algorithm. The definition of search
are: accuracy PSA is: PSA = C/R, where R is the number of
(1) Divide the target into N sub-blocks and generate corre- images returned after retrieval, and C is the number of correct
sponding sub-models. images. In the experiment, n target images are randomly
(2) Use the ICA algorithm for each sub-model to solve the obtained from each category in the target image library to be
mixed matrix [16]. retrieved as query images, and the average search accuracy
(3) Select the unmixing matrix suitable for the current of each algorithm is calculated to compare the performance
image according to the unmixing matrix in the submodel. of each algorithm. In each category, 3 target images are ran-
domly obtained for searching, and the performance of each
algorithm is compared [19].
III. ACTIVE CONTOUR IMAGE CUTTING EXPERIMENT
BASED ON IMAGE PROCESSING TECHNOLOGY C. IMAGE SEGMENTATION EXPERIMENT
A. ACTIVE CONTOUR MODEL The test environment for this experiment is Pentium CPU
The active contour (LGIF) model combines the advantages 2.90 GHz, 2 GB RAM, Windows XP 32bit, and Matlab2010a.
of the CV model and the LBF model, can effectively segment The proposed algorithm is tested through simulation experi-
images with uneven contrast and can reduce the impact of ments, and the corresponding results are given.
noise on image segmentation. However, the initial contour Regarding the evaluation method, in addition to the cal-
line C of the general model is a regular figure, such as a rect- culation of the number of iterations and calculation time,
angle, a triangle, which leads to an increase and improvement Image result of the expert manual experience segmentation
in the evolution speed of the level set. Aiming at the above are selected as the gold standard for the image of the exper-
shortcomings, a method of adding K-means clustering infor- imental simulation, namely GT (Ground Truth), and then
mation to the LGIF model is proposed. Experiments prove use different methods to obtain the segmentation results The
that this improvement effectively accelerates the convergence results were compared with it according to different evalu-
speed of the model’s active contour evolution. In order to ation criteria to obtain accurate data. This section uses the
avoid the initialization of the active contour, a penalty term following evaluation criteria:
is added to the LGIF model, and a level set function is given First, use dice similarity to measure the similarity of the
in the level set method [17]. two sets of segmentation results and the gold standard, and

19190 VOLUME 9, 2021

X. Ye, Q. Wang: Active Contour Image Segmentation Method for Training Talents

then use Jaccard similarity, which is another commonly used

standard for evaluating segmentation accuracy [20].
This experiment is divided into four parts: the experiment
compares the clustering results before and after the connected
region labeling method, the second experiment simply com-
pares the segmentation effect of the LGDF model before and
after replacing the kernel function. The third experiment ver-
ifies the effectiveness of the clustering method to obtain the
initial contour. Experiment 4 further compares the proposed
complete method combining clustering and replacing the FIGURE 1. Image cutting efficiency comparison of different models.

kernel function with the segmentation effect of the original

LGDF model.

IV. ACTIVE CONTOUR IMAGE SEGMENTATION ANALYSIS

BASED ON GRAPHICS AND IMAGE PROCESSING
TECHNOLOGY
A. IMAGE SEGMENTATION ANALYSIS OF ACTIVE
CONTOUR MODEL
From the analysis of the initial contour angle, it can be seen
that the initial contour lines of the CV model, the LBF model,
and the LGIF model are all regular circles or rectangles,
FIGURE 2. Comparison of image cutting error rates of different models.
while the initial contour lines of the K-LGIF model are closer
to the target, reducing the number of iterations. Thereby,
speeding up the image segmentation speed. Judging from the
segmentation effects in Table 1 and Figure 1, the K-LGIF is 12.5 times, 2.65 times, and 0.54 times higher than the
model is better than the CV model and the LBF model. The efficiency of CV, LBF, and LGIF models. It can be seen that
comparison of the segmentation time is shown in Table 1. K-LGI can not only ensure the segmentation effect of low-
It can be seen that compared with the CV model, LBF model, contrast images, but also it can effectively reduce the error
and LGIF model, the calculation speed of the simulation rate [22]. The experimental results show that the K-LGIF
experiment image of the K-LGIF model in this paper has model can segment the target better, but due to the complexity
been increased by 5.31 times, 3.24 times, and 2.37 times of the mouse femur image, a very small part of the trabecular
respectively [21]. bone cannot be completely segmented by the K-LGIF model.
The shortcomings of the algorithm are listed in [23].
The given K-LGIF model overcomes the shortcomings of
TABLE 1. Comparison of image segmentation efficiency of different
models. CV, LBF, and LGIF models that the initial contours of the
LGIF model are standard geometric figures. The K-LGIF
model combines the advantages of the CV model and the LBF
model, which is effective for the segmentation of Micro-CT
medicine with uneven gray levels. The slice image has a
good effect. After adding the clustering template to the LGIF
model combined with the clustering information, its initial
contour line can initially identify the target contour in the
image, which improves the running speed of the LGIF model.
The K-LGIF model also improves the segmentation effect
of the LGIF model on low-contrast images. Compared with
the CV model, LBF model, and LGIF model, the segmenta-
From the comparison of experimental segmentation tion efficiency of the K-LGIF model for simulated images
effects, neither CV nor LBF models can completely separate and Micro-CT medical images is increased by 9.12 times,
the objects in the image from the background. The LBF model 2.57 times, and 1.44 times on average, respectively. However,
is better than the CV model for image segmentation with the K-LGIF model can better improve the segmentation effect
uneven grayscale, while the LGIF and K-LGIF models. The of complex images needs further research [24].
mouse femur trabecula can be segmented more accurately,
and the latter can improve the image effect to a certain extent. B. ANALYSIS OF GRAPHIC IMAGE PROCESSING
Figure 2 indicates the comparison of error rates of different TECHNOLOGY
models. From the point of view of calculation error rate, The text-based image restoration technology uses arti-
the K-LGIF model in this paper has the least error, which ficial image description and text restoration technology.

VOLUME 9, 2021 19191

 X. Ye, Q. Wang: Active Contour Image Segmentation Method for Training Talents

However, due to its subjectivity and seriousness of the con- C. ANALYSIS OF TRAINING TALENTS FOR GRAPHICS AND
tent, it can no longer meet the requirements of today’s society IMAGE PROCESSING TECHNOLOGY
for rehabilitation methods. In addition, the content-based The computer graphics and image major is still a new research
image restoration method (CBIR technology), which starts direction and there are still shortcomings in various aspects,
from the underlying characteristics of the image, compares mainly including the following problems that need to be
the image attribute data with the database, judges the similar- resolved. First, the teaching content lacks the characteristics
ity of the image, and finds the image between the images. This of vocational education, which is the continuation of under-
can reduce human consumption, no subjective description, graduate education, and the application of talent training.
and search results are good [25]. The operation ability is relatively lacking, and the course
content is complicated and difficult for students to accept.
Secondly, insufficient attention was paid to the application
TABLE 2. Performance comparison of image algorithms in different
categories.
of technology and theoretical principles, a one-sided under-
standing, and only recognized as the expansion of secondary
vocational education. The talents cultivated by this spe-
cialty should be practical and should integrate the theory of
undergraduate education and secondary vocational education.
Comprehensive professional quality education with all-in-
one operations. Third, the content of teaching cannot keep
up with the needs of market development [28]. Computer
graphics and image technology are produced with the devel-
Table 2 is the comparison of the image retrieval method
opment of network information technology, and the rapid
based on the PSO-FWA-PAM algorithm, PSO-K-Means
development of network information technology has led to
image retrieval method, and K-means image retrieval method
the rapid development of the profession. The number and
proposed in this paper. The target image library is com-
types are diverse, and the update speed is relatively fast. Many
pared in the same environment. From the search accuracy
of the latest graphics and images technology failed to enter
of the three methods, the method proposed is superior to the
teaching in time [29]. Therefore, in view of the status quo of
two methods of PSO-K-Means image retrieval method and
this profession, efforts are required to carry out reforms and
K-means image retrieval method [26].
innovations in all aspects of the cultivation of this profession.
In order to explore the effects of the talent training model
of computer graphics and image technology processing,
the evaluation results of 100 students surveyed are shown
in Figure 4:

FIGURE 3. Comparison of image retrieval results.

Figure 3 demonstrates the three retrieval effects of the FIGURE 4. Evaluation of talent training model for graphics and image
technology processing.
PSO-FWA-PAM algorithm in the image database. The upper
left corner of the rendering is the retrieval target, and the rest
is the search feedback. It can be seen from the figure that a It can be seen from the figure that 58 students out
total of 45 retrieved images were retrieved from the African of 100 students are very satisfied with the talent training
category, and 25 related images were retrieved. Furthermore, model of computer graphics and image technology process-
the search accuracy rate was 64.5%; a total of 45 retrieved ing, accounting for 58% of the total, and they are satisfied
images were retrieved from the automobile category, and with the talent training model of graphics and image tech-
37 related images were retrieved. The accuracy rate is 78%; nology processing. It accounts for 30%. In addition, 10%
in the retrieval of dinosaurs, a total of 45 retrieved images and of students are dissatisfied with the talent training model
35 related images are obtained, and the search accuracy rate of graphics and image technology processing, and 2% of
is 72.5%. It can be seen from the experimental results that the students are very dissatisfied with the talent training model
method of the paper design has improved accuracy compared of graphics and image technology processing. Finally, there
with the original method [27]. are 2 % of students expressed a neutral attitude towards this,

19192 VOLUME 9, 2021

X. Ye, Q. Wang: Active Contour Image Segmentation Method for Training Talents

which means that they do not pay attention to or do not [3] G. Li, X. Zhao, K. Du, F. Ru, and Y. Zhang, ‘‘Recognition and evaluation
understand the talent training model of graphics and image of bridge cracks with modified active contour model and greedy search-
based support vector machine,’’ Autom. Construction, vol. 78, pp. 51–61,
technology processing. Based on the above analysis, it can Jun. 2017.
be known that most students are still satisfied with the talent [4] P. Wang and S. Wang, ‘‘Computer-aided CT image processing and mod-
training model of graphics and image technology processing, eling method for tibia microstructure,’’ Bio-Des. Manuf., vol. 3, no. 1,
pp. 71–82, Mar. 2020.
which also shows that the talent training model of graphics
[5] L. Cabaret, L. Lacassagne, and D. Etiemble, ‘‘Parallel light speed labeling:
and image technology processing will be the future develop- An efficient connected component algorithm for labeling and analysis
ment trend [30]. on multi-core processors,’’ J. Real-Time Image Process., vol. 15, no. 1,
pp. 173–196, Jun. 2018.
[6] A. Ebrahimi and G. B. Loghmani, ‘‘B-spline curve fitting by diagonal
V. CONCLUSION approximation BFGS methods,’’ Iranian J. Sci. Technol., Trans. A, Sci.,
After explaining the importance of image segmentation and vol. 43, no. 3, pp. 947–958, Jun. 2019.
its main purpose, this article reviews two types of image [7] Q. Wang, K. Wu, X. Wang, Y. Sun, X. Yang, and X. Lou, ‘‘Automatic
detection and classification of foreign bodies of dumplings based on X-
segmentation methods, including edge-based image segmen- ray,’’ J. Comput.-Aided Des. Comput. Graph., vol. 30, no. 12, p. 2242,
tation methods and region-based image segmentation meth- 2018.
ods, and introduces various types of typical segmentation [8] H. Zhao, G. W. Bryant, W. Griffin, J. E. Terrill, and J. Chen, ‘‘Validation
of SplitVectors encoding for quantitative visualization of large-magnitude-
methods. Then, this study focuses on the methods of the range vector fields,’’ IEEE Trans. Vis. Comput. Graphics, vol. 23, no. 6,
active contour model, including the parametric active contour pp. 1691–1705, Jun. 2017.
model and geometric active contour model. Among them, [9] H. Sarfraz, A. A. G. Hassan, and A. Sohail, ‘‘Talent development &
the Snake model, CV model, LBF model, LGDF model, excellence the reflection of exchange rate exposure and working capital
management on manufacturing firms of Pakistan,’’ Talent Develop. Excel-
etc. are explained and compared with the advantages and lence, vol. 12, no. 2s, pp. 684–698, 2020.
disadvantages of related methods. [10] C. E. Thomas, G. Abbott, P. B. Gastin, and L. C. Main, ‘‘Construct validity
Aiming at the shortcomings of traditional GVF models and reliability of the talent development environment questionnaire in
caribbean youth track and field athletes,’’ PLoS ONE, vol. 15, no. 1,
that are difficult to segment complex images, this paper pro- Jan. 2020, Art. no. e0227815.
poses an active contour model algorithm based on saliency. [11] A. A. Mohammed, A. Hafeez-Baig, and R. Gururajan, ‘‘A qualitative
The algorithm mainly uses the gradient descent flow vec- research to explore practices that are utilised for managing talent devel-
tor field generated by the edge map of the image saliency opment in the higher education environment: A case study in six australian
universities,’’ J. Industry-Univ. Collaboration, vol. 1, no. 1, pp. 24–37,
map to perform a primary evolution of the active contour, Feb. 2019.
and then uses the mixed gradient descent flow generated [12] S. Hartati, D. Safitri, and A. Marini, ‘‘Talent development & excellence
by the edge of the original image and the saliency map to bullying behavior in early childhood: Study at early childhood educa-
tion institution in East Jakarta in Indonesia,’’ Talent Develop. Excellence,
perform the secondary evolution of the active contour. The vol. 12, nos. 1, pp. 55–63, 2020.
first evolution can make the contour line quickly approach [13] X. Pan, Y. Zhou, F. Li, and C. Zhang, ‘‘Superpixels of RGB-D images for
the edge of the target we are interested in in the image, indoor scenes based on weighted geodesic driven metric,’’ IEEE Trans. Vis.
Comput. Graphics, vol. 23, no. 10, pp. 2342–2356, Oct. 2017.
and the second evolution can make the contour line get the
[14] G. Wang, J. Lu, Z. Pan, and Q. Miao, ‘‘Color texture segmentation based
precise target edge. Experiments illustrate that the algorithm on active contour model with multichannel nonlocal and tikhonov reg-
in this paper can effectively improve the ability of the gra- ularization,’’ Multimedia Tools Appl., vol. 76, no. 22, pp. 24515–24526,
dient descent flow active contour model to segment complex Nov. 2017.
[15] X. Liao, Z. Yuan, Q. Tong, J. Zhao, and Q. Wang, ‘‘Adaptive
images. localised region and edge-based active contour model using shape con-
This article describes the advantages of different active straint and sub-global information for uterine fibroid segmentation in
contour models. Algorithms that achieve good results in one ultrasound-guided HIFU therapy,’’ IET Image Process., vol. 11, no. 12,
pp. 1142–1151, Dec. 2017.
type of problem may not work well in another type of prob-
[16] B. Han and Y. Wu, ‘‘SAR river image segmentation by active contour
lem. Facing a new segmentation task, it is often necessary model inspired by exponential cross entropy,’’ J. Indian Soc. Remote Sens.,
to use a trial method to determine which algorithm to use, vol. 47, no. 2, pp. 201–212, Feb. 2019.
and there is no general solution. Most of the prior informa- [17] R. Jin and G. Weng, ‘‘A robust active contour model driven by pre-fitting
bias correction and optimized fuzzy c-means algorithm for fast image
tion used by current active contour models are only prior segmentation,’’ Neurocomputing, vol. 359, pp. 408–419, Sep. 2019.
to contours. However, in practical problems, more available [18] Q. Zheng, H. Li, B. Fan, S. Wu, J. Xu, and Z. Cao, ‘‘Modified localized
prior information (such as prior texture, prior color, etc.) are multiplicative graph cuts based active contour model for object segmen-
tation based on dynamic narrow band scheme,’’ Biomed. Signal Process.
required and it is not convenient to obtain prior contours. Control, vol. 33, pp. 119–131, Mar. 2017.
Therefore, the key issue to obtain and use prior information [19] S. Niu, N. Li, Z. Guo, R. Wang, Y. Guo, L. Wu, and J. Zhao, ‘‘Robust
in the active contour model more efficiently. boundary extraction of great lakes by blocking active contour model using
Chinese GF-3 SAR data: A case study of Danjiangkou reservoir, China,’’
J. Eng., vol. 2019, no. 20, pp. 6876–6879, Oct. 2019.
REFERENCES [20] W. Zhou, Y. Yi, Y. Gao, and J. Dai, ‘‘Optic disc and cup segmentation in
[1] S. P. Khadilkar, S. R. Das, and M. H. Assaf, ‘‘Face identification based retinal images for glaucoma diagnosis by locally statistical active contour
on discrete wavelet transform and neural networks,’’ Int. J. Image Graph., model with structure prior,’’ Comput. Math. Methods Med., vol. 2019,
vol. 19, no. 4, pp. 634–654, 2019. no. 2, pp. 1–16, 2019.
[2] G. Wu, B. Masia, A. Jarabo, Y. Zhang, L. Wang, Q. Dai, T. Chai, and Y. Liu, [21] Y. Wu, X. Liu, D. Zhou, and Y. Liu, ‘‘Adaptive active contour model driven
‘‘Light field image processing: An overview,’’ IEEE J. Sel. Topics Signal by image data field for image segmentation with flexible initialization,’’
Process., vol. 11, no. 7, pp. 926–954, Oct. 2017. Multimedia Tools Appl., vol. 78, no. 1, pp. 1–26, 2019.

VOLUME 9, 2021 19193

 X. Ye, Q. Wang: Active Contour Image Segmentation Method for Training Talents

[22] A. R. Raju, S. Pabboju, and R. R. Rao, ‘‘Hybrid active contour model XINGHUO YE was born in Ezhou, Hubei, China,
and deep belief network based approach for brain tumor segmentation and in 1982. He received the Ph.D. degree from Wuhan
classification,’’ Sensor Rev., vol. 39, no. 4, pp. 473–487, Jul. 2019. University, China. He currently works with the
[23] Z. Zhu, Y. Tang, J. Hu, and M. An, ‘‘Coastline extraction from high- School of Information Engineering, Huzhou Uni-
resolution multispectral images by integrating prior edge information with versity. His research interests include data mining
active contour model,’’ IEEE J. Sel. Topics Appl. Earth Observ. Remote and bioinformatics.
Sens., vol. 12, no. 10, pp. 4099–4109, Oct. 2019.
[24] G. Liu, Y. Dong, M. Deng, and Y. Liu, ‘‘Magnetostatic active contour
model with classification method of sparse representation,’’ J. Electr.
Comput. Eng., vol. 2020, pp. 1–10, Jul. 2020.
[25] H. Chen, X. S. Yu, and C. D. Wu, ‘‘Fast image segmentation algorithm
based on parametric level set active contour model,’’ Dongbei Daxue
Xuebao/J. Northeastern Univ., vol. 40, no. 1, pp. 6–10, 2019.
[26] B. Wang, X. Gao, D. Tao, and X. Li, ‘‘A nonlinear adaptive level set for
image segmentation,’’ IEEE Trans. Cybern., vol. 44, no. 3, pp. 418–428,
Mar. 2014.
[27] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille,
‘‘DeepLab: Semantic image segmentation with deep convolutional nets,
atrous convolution, and fully connected CRFs,’’ IEEE Trans. Pattern Anal.
QIANYI WANG was born in Linyi, Shandong,
Mach. Intell., vol. 40, no. 4, pp. 834–848, Apr. 2018.
[28] J. Huang, X. Yang, and Y. Chen, ‘‘A fast algorithm for global minimization China, in 1994. She received the master’s
of maximum likelihood based on ultrasound image segmentation,’’ Inverse degree from Northeast Forestry University, China.
Problems Imag., vol. 5, no. 3, pp. 645–657, 2011. She currently works with the AI Lab Artifi-
[29] S. Bargoti and J. P. Underwood, ‘‘Image segmentation for fruit detection cal Intelligence Laboratory, Beijing Chengshi
and yield estimation in apple orchards,’’ J. Field Robot., vol. 34, no. 6, Wanglin Information Technology Company Ltd.
pp. 1039–1060, Sep. 2017. Her research interests include image processing,
[30] V. Singh and A. K. Misra, ‘‘Detection of plant leaf diseases using image speech recognition, and information security.
segmentation and soft computing techniques,’’ Inf. Process. Agricult.,
vol. 4, no. 1, pp. 41–49, Mar. 2017.

19194 VOLUME 9, 2021