2021 International Conference on Networking, Communications and Information Technology (NetCIT)

Computer network connection enhancement

2021 International Conference on Networking, Communications and Information Technology (NetCIT) | 978-1-6654-0070-1/21/$31.00 ©2021 IEEE | DOI: 10.1109/NetCIT54147.2021.00063

optimization algorithm based on convolutional

neural network
Hanju Li
Sichuan University, Chengdu, 610225, China
[email protected]
Corresponding author

Abstract—Computer network topological structure is a interactivity of information, thereby expanding the

kind of physical composition mode, which is composed of three topological structure of the computer network [5].
aspects, namely, network computers or network equipment,
nodes and lines in network transmission media. > This type of II. ALGORITHM PRINCIPLE OF CONVOLUTIONAL NEURAL
optimization problem is often solved when expanding the NETWORK
topology of a computer network. This paper establishes a
reasonable network structure model based on the
The artificial neural network system is a pioneering work
convolutional neural network to express the complex nonlinear in the middle of the last century. It is widely used and has
relationship in the computer network connection, and proposes good parallel processing, distributed information storage,
an optimization algorithm based on the characteristics of the self-organizing and self-learning capabilities. Convolutional
computer network. I hope to provide some reference and neural networks have been widely used in many fields,
reference for promoting the use of convolutional neural especially It is more concerned by researchers in the fields of
networks in computer networks. information processing, pattern recognition, intelligent
control, and system modeling [6]. The convolutional neural
Keywords—Convolutional neural network algorithm, network algorithm is a supervised learning algorithm that
computer network topology, network characteristics uses learning samples to use backpropagation methods. The
deviation of the network is adjusted repeatedly, so that the
I. INTRODUCTION output structure and the expected vector are infinitely close
The mode of computer network connection also refers to [7].
the computer network topology, which is the node formed
The typical structure of convolutional neural network
between the computer and the network, as well as the formed
usually includes: input layer, convolution layer,
line [1]. The nodes of the network are mainly composed of
down-sampling layer (also known as pooling layer), fully
two types: the nodes that convert and exchange information,
connected layer and output layer [8], as shown in Figure 1.
including some devices such as switches, hubs, and
controllers; and the access nodes, including the computer
host and interrupt. The line connected between the computer
and the network is a variety of transmission media including
tangible and intangible [2].
With the rapid development of computer network
technology, the computer network topology is becoming
more and more complex, so the expansion of its structure is Fig.1 Typical structure of convolutional neural network
more important. At present, it is more important to choose
the right node and improve the capacity and efficiency of the Due to the deep structure of the convolutional neural
network. The sub-problem is the problem of connection network, these features are conducted layer by layer through
enhancement [3]. It is very important to add additional the layers of the network, so that the expression of the
connections without adding additional nodes, and it is also a low-level simple features of the computer network
key to improving the cost-effectiveness of the network. In connection to the high-level abstract features can be obtained,
the current situation, there are not many ways to achieve which has a stronger discrimination than the traditional
computer network connection enhancement, but to maximize artificial design features And generalization performance.
profits under the right connection, the best way is to optimize However, due to the complexity and diversity of the
the network algorithm [4]. computer vision field, there are still many problems that
need to be solved urgently in the in-depth study of
This article starts with the structure of the network and
convolutional neural networks in specific fields [9]. For
the topology of the network. For example, in the selection of
example, in the application of computer network connection
network nodes, select appropriate and reasonable nodes, and
classification to reduce the over-fitting problem of the
then add these nodes to the running computer network. At
network to improve the discriminative performance of the
the same time, it is necessary to ensure that the work of other
convolutional neural network; in the scene labeling
nodes is not affected. It can achieve the purpose of
application, the convolutional neural network is used to
improving the connection efficiency of the network and the
improve the accuracy of the labeling while improving the
network bandwidth capacity, and it also strengthens the

978-1-6654-0070-1/21/$31.00 ©2021 IEEE 281

DOI 10.1109/NetCIT54147.2021.00063
Authorized licensed use limited to: NUST School of Electrical Engineering and Computer Science (SEECS). Downloaded on November 28,2024 at 08:46:20 UTC from IEEE Xplore. Restrictions apply.
 result of the scene labeling. Regional consistency [10]. gradient explosion or disappearance, which makes deep
learning models difficult to train. In order to solve the above
III. OPTIMIZED COMPUTER NETWORK CONNECTION problems, this part will introduce short-circuit jump links,
ENHANCEMENT ALGORITHM BASED ON CONVOLUTIONAL and its structure is shown in Figure 2.
NEURAL NETWORK

A. Network model construction

The structure of the network model determines the data
fitting ability and generalization performance of the network
model. In actual applications, the structure definition of the
network model needs to define the depth of the network
according to the amount of training data and the
characteristics of the data set itself. , The hierarchical
functions of the network, and the hyperparameters in the
network. As a heuristic method, the standard of
convolutional neural network structure definition mainly
comes from some experience summaries. In the actual
experiment process, it is usually necessary to debug the
structure and hyperparameters of the network based on the
training data set to obtain the best network performance [11].
The numerical characteristics of each network into the
network model data set have different dimensions. The data
of different dimensions will increase the training time of the
convolutional neural network. Therefore, in order to speed Fig.2 Short-circuit jump link structure diagram
up the convergence of the training convolutional neural
network, this paper uses minimum-maximum normalization Assuming that the computer network of the f-th layer of
to linearly transform the original data, and map the data the neural network has undergone data preprocessing and
values to [0,1]. Assuming that the input numerical features data conversion, the feature map of the image is m, and then
remain unchanged, the formula is as shown in (1) the features of the network convolutional neural network
data are extracted through a layer of convolution operation.
xi  min X In order to prevent the gradient dispersion phenomenon
xi generated by the neural network during the training process,
max X  min X (1) this paper will batch normalize the output after the previous
convolution operation, so that the mean and variance of the
Among them, max(X) and min(X) are respectively output result are 0 and 1, respectively.
expressed as the maximum value and the minimum value in B. Computer network model fusion based on convolutional
the numerical xin feature. The forward conduction process of neural network
the network model is the expression process of the input data
The diversity of the network comes from the structural
at various levels in the network. The forward conduction of
differences of each convolutional neural network. For the
the network is based on the training results of a specific
same input network model, each network will output a
training set. A new feature expression of the input image will
prediction based on the input. The predictions from multiple
be output at each level of the network. For the type-specific
networks need to be fused in an appropriate way. The
features of the network into the network model, this article
convolution network fusion model is shown in Figure 3.
will use one-hot encoding or dumb Variable coding is used
for numerical operation. Assuming that the number of
numerical features is n, the number of dimensions of
categorical features after one-hot encoding or dummy
variable encoding is n. Then the number of dimensions of the
total combined feature N2, Nęz is solved to satisfy the
following constraints (2)(3):

( N  1) 2  (nm  nc )  0 (2)

N 2  (nm  nc ) t 0 (3) Fig.3 Convolutional network fusion model

Based on the convolution, our fusion model introduces

The features of the last layer of the network are usually the function of the mean of the random variable instead of
converted into the probability distribution output of the the mean of the function of the random variable. As shown
category through the full connection. In addition, the in formula (4)(5),
high-level feature expression of the convolutional neural
network has strong discrimination and generalization 1 § Ui ·
performance, and can be used in the application of transfer Si [1  tanh ¨ ¸
learning [12]. Convolutional networks have the problem of 2 © T ¹ (4)

282

Authorized licensed use limited to: NUST School of Electrical Engineering and Computer Science (SEECS). Downloaded on November 28,2024 at 08:46:20 UTC from IEEE Xplore. Restrictions apply.
 Ui ¦WiSi  Ii ¦Wi Si  Ii (5)
prediction results with better performance. Therefore, it is
necessary to consider a more flexible network connection
enhancement mechanism.
Among them, (x) represents the mean value of the
This article uses pseudo-variable encoding for one of the
variable x. Substituting the mean field variable Vi for the
features, and uses one-hot encoding for the other categorical
random variable mean (Si), the computer network model
features. After the preprocessing step, the number of feature
convolution kernel has an impact on the model in two
dimensions has been expanded from 41 to 121. After that,
aspects. On the one hand, the size of the convolution kernel
we convert each data of the convolutional neural network
determines the scale of the network model, and also affects
with 121-dimensional features into a 1 1×1 1×1 array. The
the size of the output feature network; on the other hand, the
operation is shown in Figure 4.
number of convolution kernels also affects the number of
feature networks [13]. However, for a specific data set, there
is currently no effective method to determine the optimal
convolution kernel size and the number of feature networks.
Therefore, this chapter builds multi-column
convolutional neural networks with different structures by
setting different convolution kernel sizes and the number of
feature networks, and trains each convolutional neural
network model on the training set separately to achieve
different network models The purpose of structure fitting
training data set. Finally, by fusing the output results of each Fig.4 Data conversion operation
network, a relatively good prediction result is obtained.
C. Logistics connection optimization based on convolutional The main reason for this result is that convolutional
neural network neural networks are superior to machine learning methods in
terms of feature extraction capabilities. The accuracy of
After the convolutional neural network model fusion, the traditional machine learning methods is lower than that
convolutional neural network has excellent discrimination based on convolutional neural networks. This is because the
ability for the extraction of complex feature networks. representation of normal and abnormal data after feature
Overfitting is a common problem in the field of machine extraction is similar in terms of low-level features. However,
learning. The main reason is that the learning model overfits convolutional neural networks can automatically extract
the features of the training set, which leads to the lack of complex high-level features from these similar low-level
generalization performance of the model and affects the features.
feature extraction and judgment of the test data set. . The
input image after different preprocessing has a richer For the multi-scale salient regions (Q1, Q2, Q3) extracted
expression, which is helpful for the network to determine the by the computer network, it is necessary to use the
final image category. Finally, the model's judgment of the convolutional neural network trained on the large-scale
object category is obtained based on the mean value of the network data set to extract the features U(K) respectively.
multi-column neural network prediction. As shown in These parameters are passed in the large-scale network The
formula (6), training on the data set changes from a random state to a
stable convergence state. Finally, the characteristic
expression of the multi-scale significant area of the entire
1 N
MCDNNi ¦ DNN j
N j1
computer network is obtained (7):
(6)
U U 1  U 2  U 3 (7)
MCDNN represents the probability of the i-th category
output by the multi-column convolutional neural network. The accuracy of computer network topology shows the
DNN represents the number of convolutional neural proportion of all correctly labeled computer network
networks; j represents the index of the j-th network in the topologies in the data set occupying the computer network
multi-column convolutional neural network. topology of the entire data set. This evaluation method
First of all, in the choice of network structure, the use of reflects the accuracy of scene labeling to a certain extent, but
exactly the same convolutional neural network structure may ignores an important feature of a data set, that is, the number
limit the difference of the network for feature learning to a of computer network topologies of each category label in the
certain extent. The advantage of the multi-column network data set is not consistent [15]. This characteristic causes the
structure compared to the single-column network is that evaluation method of computer network topology accuracy
different networks can learn the characteristics of the input to be affected by the large number of computer network
image from different angles. Then, increasing the difference topology categories. Therefore, as a supplement to the
between different networks through changes in the network evaluation method of computer network topology accuracy,
structure is a potential method to improve the overall the category average accuracy evaluation method has also
performance of the network. been widely adopted. The category average accuracy
evaluation method first obtains the computer network
Secondly, the method of connecting and enhancing the topology accuracy of each category itself, and then sums and
prediction of multi-column networks by simply taking the averages the accuracy of each category to obtain the final
average value is relatively simple, because the discriminant scene labeling accuracy. The category average accuracy
performance of different networks is different [14]. Network evaluation method ensures that the computer network

283

Authorized licensed use limited to: NUST School of Electrical Engineering and Computer Science (SEECS). Downloaded on November 28,2024 at 08:46:20 UTC from IEEE Xplore. Restrictions apply.
 topology accuracy of each category has the same influence enhancement optimization algorithms, such as simulated
weight on the final scene labeling accuracy. This relevance annealing algorithms and genetic algorithms. This article
means that under the existing convolutional neural network only discusses the convolutional neural network algorithm
framework, by improving the characteristics of general alone. Compared with other algorithms, whether it is more
convolutional neural networks, it is beneficial to further feasible, this article does not do more in-depth research. This
improve the accuracy of computer network topology results. is also the future research direction of computer network
connection enhancement optimization algorithm based on
IV. CONCLUSION convolutional neural network.
This article is based on the research of convolutional
neural network, mainly focusing on the key technology of REFERENCES
computer network connection enhancement. Aiming at the [1] Liu Yu, Liu Houquan. Facial image restoration based on adversarial
training and convolutional neural network[J]. Computer Engineering
problem of object recognition, this paper takes the and Applications, vol.055, no.002, pp.110-115,136, 2019.
characteristics of the network as the research of the structure
[2] Chen Z, Hong Y, Wang J, et al. Monocular visual odometer based on
of the convolutional neural network model as an entry point. cyclic convolutional neural network [J]. Jiqiren/Robot, vol.41,
By fusing multiple columns of convolutional neural no.2,pp.147-155, 2019 .
networks with different structures, objects with better [3] Zhang Zhenhuan, Zhou Cailan, Liang Yuan. Optimized
stability than a single network structure are obtained. Convolutional Neural Network Garment Classification Algorithm
Recognition effect. No matter which algorithm is adopted to Based on Residual Error[J]. Computer Engineering and Science,
simulate the network connection, it is necessary to evaluate vol.040, no.002, pp.354-360, 2018.
its performance in the end. This paper selects the [4] Jia Kai, Duan Xintao, Li Baoxia, et al. Image super-resolution
enhancement algorithm based on dual-channel convolutional neural
convolutional neural network algorithm. According to the network [J]. Computer Applications, vol.038, no.012, pp.3563-3569,
analysis of the calculation process, it is concluded that the 2018.
algorithm that can make the network connection achieve [5] Tu Wenbo, Yuan Zhenming, Yu Kai. Chinese word segmentation
better results in the convolutional neural network algorithm method of convolutional neural network without pooling layer[J].
is the average field neural network algorithm, because it is Computer Engineering and Applications, vol.17,no.11,pp.21-22,2020.
not only fast but also The result is closer to the real situation, [6] Ouyang Ning, Zeng Mengping, Lin Leping. Super-resolution
so the algorithm is feasible in terms of enhancing and reconstruction based on parallel convolutional neural network[J].
Computer Applications, vol.37,no.4,pp.52-53,2017.
optimizing the computer network connection and expanding
the topology of the network structure. [7] GAO Lei, FAN Bing-Bing, HUANG Sui. Improved Convolutional
Neural Network Image Classification Algorithm Based on Residual
According to the current research situation, there are still Error [J]. Computer System Applications, vol.028, no.007,
pp.139-144, 2019.
many ways to enhance and optimize computer network
[8] Sun Yuechi, Ping Wei, Xu Minglei. Human behavior recognition
connection efficiency and increase network communication based on optimized convolutional neural network structure[J].
capacity, but these methods and measures have more or less Computer Applications and Software, vol.38,no.2,pp.8,2021.
problems of this kind, or they are not practical. And the [9] Zhu Fuli, Yang Lei, Ji Bo. Face detection method based on enhanced
economy, and the enhanced and optimized computer parallel cascaded convolutional neural network[J]. Computer
network connection ground averaging neural network Applications and Software, vol.37,no.11,pp.61-62,2020.
algorithm proposed in this paper can well meet the above [10] Wen Guangzhao, Xu Shinan, Ma Yunhe, et al. Expression recognition
requirements and has good economy [16]. algorithm based on binary dense convolutional neural network [J].
Computer and Digital Engineering, vol. 48, no. 3, pp. 62-63, 2020 .
By establishing a scientific and reasonable field average [11] Lu Feiyu. Traffic sign detection and recognition algorithm based on
neural network model, it is helpful to carry out computer improved convolutional neural network[J]. Industrial Control
network connection enhancement and optimization in the Computer, vol. 32,no.6,pp.3-4,2019.
field average neural network algorithm research work, and [12] Gao Youwen, Zhou Benjun, Hu Xiaofei. Research on Convolutional
then judge the network effect. It should be noted that when Neural Network Image Recognition Based on Data Augmentation[J].
Computer Technology and Development, vol.28,no.8,pp.4-5,2018.
using the function method to construct a model, we should
[13] He Yufei, Gao Hongwei. Single-frame image super-resolution
strengthen the problem of constructing the objective function. reconstruction based on multilayer connected convolutional neural
Based on the enhanced optimization network connection network[J]. Computer Applications and Software, vol.36,no.5,
average field neural network algorithm, in order to analyze pp.6-7,2019.
this neural network algorithm in depth and achieve the [14] Long Min, Tong Yueyang. Research on face live detection algorithm
desired result, the first step is to construct an excellent using convolutional neural network[J]. Computer Science and
average field neural network model. The construction of the Exploration, vol.12,no.10,pp.13-15,2018.
network algorithm model needs to pay attention to many [15] Zheng Chenghao, Liu Bing, Zhou Yong. Scale-independent cascaded
factors, especially in the process of constructing and convolutional neural network face detection algorithm [J]. Computer
Application Research, vol. 036, no.002, pp.593-597,605, 2019.
selecting the objective function to choose appropriate,
[16] Zhang Wenda, Xu Yuelei, Ni Jiacheng, et al. Image target recognition
scientific and effective. Based on the characteristics of algorithm based on multi-scale block convolutional neural network
computer networks, there are other different connection [J]. Computer Applications, vol.036, no.004, pp.1033-1038, 2016.

284

Authorized licensed use limited to: NUST School of Electrical Engineering and Computer Science (SEECS). Downloaded on November 28,2024 at 08:46:20 UTC from IEEE Xplore. Restrictions apply.