Convolutional Neural Networks
Quan Zhang
School Of Electronic & Information Engineering, Suzhou University of Science and Technology, Suzhou 215009, China
Keywords: Convolutional Neural Network, feature extraction, deep learning.
Abstract: The earliest Convolution Neural Network (CNN) model is leNet-5 model proposed by LeCun in 1998.
However, in the next few years, the development of CNN had been almost stopped until the article
‘Reducing the dimensionality of data with neural networks’ presented by Hinton in 2006. CNN started
entering a period of rapid development. AlexNet won the championship in the image classification contest
of ImageNet with the huge superiority of 11% beyond the second place in 2012, and the proposal of
DeepFace and DeepID, as two relatively successful models for high-performance face recognition and
authentication in 2014, marking the important position of CNN. Convolution Neural Network (CNN) is an
efficient recognition algorithm widely used in image recognition and other fields in recent years. That the
core features of CNN include local field, shared weights and pooling greatly reducing the parameters, as
well as simple structure, make CNN become an academic focus. In this paper, the Convolution Neural
Network’s history and structure are summarized. And then several areas of Convolutional Neural Network
applications are enumerated. At last, some new insights for the future research of CNN are presented.
1 INTRODUCTION
Convolutional Neural Network is a category of
neural networks that have proved very effective in
areas such as image recognition and classification
(Krizhevsky, 2012). Convolutional Neural Network
has made a series of breakthrough research results in
the fields of face recognition (Lawrence, 1997),
sentence classification (Kim, 2014), semantic
segmentation and so on. So, it is of great value to
review the works in this research field.
Convolutional Neural Network is inspired by the
biology of the visual cortex. A small percentage of
cells in the visual cortex are sensitive to the visual
area of a particular part. An interesting experiment
by Hubel and Wiesel (Kandel, 2009) in 1962
illustrates this concept in detail. They demonstrated
that some individual neurons in the brain respond
only when they are in the edges of a particular
direction. For example, some neurons are only
excited about vertical edges, while others are excited
about horizontal or diagonal edges. Hubel and
Wiesel found that all of these neurons are arranged
in the form of columnar structures and work together
to produce visual perception. The idea that a
particular component of such a system has a
particular task (neuronal cells in the visual cortex
looking for specific features) is equally applicable in
machines, which is the basis of the CNN.
Convolution has important points that local
connection and weight sharing. Local connection
and weight sharing reduce the amount of parameters,
greatly reducing training complexity and alleviating
over-fitting. At the same time, weight sharing also
gives the convolutional network tolerance for
translation.
After years of development, the Convolution
Neural Network has been gradually extended to
more complicated fields such as pedestrian
detection, behavior recognition and posture
recognition from the initial simpler handwritten
character recognition, etc. Recently, the application
of Convolutional Neural Networks further develops
to a deeper level of artificial intelligence, such as
natural language processing, speech recognition, and
others. Recently, Alpha go, an artificial intelligence
go program developed by Google, has successfully
used convolutional neural network to analyze the
information of Go board and win the European
Champion and Go champion successively in the
Challenge, attracting wide attention. From the
current trend of research, since the application
prospect of convolutional neural network is full of
possibilities, it also faces some research problems,
434
Zhang, Q.
Convolutional Neural Networks.
In 3rd International Conference on Electromechanical Control Technology and Transportation (ICECTT 2018), pages 434-439
ISBN: 978-989-758-312-4
Copyright © 2018 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

such as how to improve the structure of Convolution
Neural Network to increase the learning ability, and
how to convert Convolution Neural Network from a
reasonable form into the new application model.
The outline of the paper is as follows. Section
two describes the background to the CNN including
the theoretical proposition stage, the model
realization stage, and the extensive research stage.
Section three defines the structure of CNN that
mainly composed of input layer, convolution layer,
subsampling layer (pooling layer), fully connected
layer and output layer. Section four gives two
examples of Convolution Neural Network
applications and discusses the future of other
applications.
2 BACKGROUND
The history of Convolutional Neural Networks can
be roughly divided into three stages: the theoretical
proposition stage (Yandong, 2016), the model
realization stage (Fukushima, 1988), and the
extensive research stage.
2.1 The theoretical proposition stage
In the 1960s, Hubel showed that the biological
information from the retina to the brain is stimulated
by multiple levels of receptive field. In 1980, for the
first time, Fukushima proposed Neocognitron
(Ballester, 2016) based on the theory of receptive
fields. Neocognitron is a self-organizing multi-layer
neural network model. The response of each layer is
obtained by the local sensory field of the upper
layer. The recognition of the model is not affected
by position, small shape changes and the size of the
scale. Unsupervised learning by Neocognitron is
also the dominant learning method in early studies
of convolutional neural networks
2.2 The model realization stage
In 1998, Lecun proposed LeNet-5 to use a gradient-
based backpropagation algorithm to supervise the
network. The trained network converts the original
image into a series of feature maps through the
convolutional layer and the down-sampling layer
alternately connected. Finally, the feature
representation of the images is classified by the fully
connected neural network. Convolutional kernel
completed the receptive field function, and it can
lower the local area information through the
convolution kernel excitation to a higher level. The
successful application of LeNet-5 in the field of
handwritten character recognition has drawn the
attention of academia to the convolutional neural
network. In the same period, research on
Convolutional Neural Networks in speech
recognition, object detection, face recognition and so
on has been gradually carried out.
2.3 The extensive research stage
In 2012, AlexNet proposed by Krizhevsky won the
championship in the image classification contest of
ImageNet, a large image database, with the huge
superiority of 11% beyond the second place, making
the Convolutional Neural Network become an
academic focus. After AlexNet, new models of
convolutional neural networks have been proposed,
such as Visual Geometry Group (VGG), Google's
GoogLeNet, Microsoft's ResNet, etc. These
networks refresh AlexNet Record created on
ImageNet. Furthermore, convolutional neural
network is continuously merged with some
traditional algorithms, and with the introduction of
migration learning method, the application of
Convolutional Neural Networks has been rapidly
expanded. Some typical applications include:
Convolutional neural network combined with
Recurrent Neural Network (RNN) for image
summarization and image content quiz; Convolution
Neural Networks have achieved significant accuracy
gains in small sample image recognition databases;
and video-oriented behavioral recognition models -
3D Convolutional Neural Networks.
3 STRUCTURE
As shown in the figure 1, the typical Convolutional
Neural Network is mainly composed of input layer,
convolution layer, subsampling layer (pooling
layer), fully connected layer and output layer.
Convolutional Neural Networks
435

Figure 1: The structure of CNN [12]
3.1 Input layer
When we enter an image, what the computer sees
is a matrix of pixel values. Depending on the
resolution and size of the image, the computer will
see different matrices, such as a 32 x 32 x 3 matrix
(3 refers to RGB values). Each digit in the matrix
has a value from 0 to 255, which describes the
pixel's gray level at that point.
In essence, each picture can be expressed as a
matrix of pixel values.
3.2 Convolution layer
The fundamental purpose of convolution is to
extract features from the input picture.
Convolution uses a small square matrix, which
preserves the spatial relationships among pixels, to
learn image features.
The convolution operation is shown in the
figure 2. The matrix that slides the filter on the
original image and performs the convolution
operation is called the feature map.
Figure 2: Convolution operation (slide=1) [13]
For every feature map, all neurons share the
same weight parameter that is known as filter or
kernel (yellow part in the figure 2). The filter is a
feature detector for the original input picture.
Different filters will produce different feature
maps for the same picture. By simply adjusting the
filter values, we can perform effects such as edge
detection, sharpening, blurring, etc. that mean
different filters detect different features, such as
edges, curves, etc. from the picture.
The image specifications continue to decrease
when the step size is increased. Filling the ‘0’
borders around the input image can solve this
problem.
As shown in the figure 3, the original shape of
the image is preserved after adding ‘0’ to the
image. This is called same padding because the
output image is as the same size as the input
image.
The size of the image has also been preserved
due to the ‘0’ borders.
ICECTT 2018 - 3rd International Conference on Electromechanical Control Technology and Transportation
436

Figure 3: Same padding (slide=2) [13]
3.3 Subsampling layer
The subsampling layer is also called the pooling
layer.
Sometimes the image is too large and we need to
reduce the number of training parameters .It is
required to periodically introduce a pooling layer
between subsequent convolution layers. The only
purpose of pooling is to reduce the size of the image
space.
Pooling can take many forms: Max Pooling,
Average Pooling, Sum Pooling, and so on. The most
common form of pooling is Max Pooling that is
shown in the figure 4.
Figure 4: Max pooling [14]
Pooling operations are applied to each feature
map separately
After convoluting and pooling, the image still
retains most of the information as well as the size of
the image has been reduced. The main role of the
Pooling layer is subsampling, which further reduces
the number of parameters by removing unimportant
samples in the Feature Map.
3.4 Fully connected layer
"Fully connected" means that every neuron in the
upper level is interconnected with every neuron in
the next level.
Convolution and pooling layer is a feature
extractor, the fully connected layer is a classifier.
The fully connected layer receives the output of
the upper layer (which represents the feature map of
the higher-level features) and determines which
category these features best fit.
For example, if the program determines the
content of a picture as a dog, the feature map with
higher values represent some of the more advanced
features such as claws or four legs. Similarly, if the
program determines the content of a picture as a
bird, the feature map with higher values represent
some of the more advanced features such as wings
or beaks.
The fully connected layer is the Multi-Layer
Perceptron that uses the Softmax excitation function
as the output layer. The Softmax function converts a
vector of any real value into a vector of elements 0-1
and 1.
In general, a fully connected layer observes
which category the advanced features most closely
matches and what weight they have. When
calculating the weight and the dot product between
previous layers, we can get the correct probability
for the different categories.
This layer process input, and then output an N-
dimensional vector (N represents the number of
categories). Each number of N-dimensional vector
represents the probability of a particular category.
Convolutional Neural Networks
437

3.5 Output layer
The fully connected layer is the Multi-Layer
Perceptron that uses the Softmax excitation function
as the output layer.
4 APPLICATION
CNN plays a very important role in different fields,
such as image processing, natural language
processing, etc. The next few cases are important
applications based on CNN.
4.1 Instance
4.1.1 Traffic Statistics
The traffic is modeled as images, and traffic
classification tasks become image classification.
Traffic statistics takes pedestrian counting for an
example. The task of pedestrian counting uses video
image analysis techniques to automatically tally the
number of pedestrians passing a scene over a certain
period of time.
Firstly, rough human head detection is carried
out based on the Adaboost human head detector.
Secondly, a large number of false targets are
eliminated by using the CNN-SVM human head
classifier through migration learning technology to
ensure a high detection accuracy rate. CNN, as a
feature extractor, is used to train linear SVM
classifier.
Finally, according to the head information, the
target of the head of the related person is found
accurately by the method of region restriction and
feature matching, which greatly improves the
counting accuracy.
Traffic statistics can be used in a wide range of
areas such as commerce, finance, hospitality and
transportation because traffic statistics is an
important piece of information in many industries.
4.1.2 Medical image segmentation
In the traditional medical image processing
techniques and machine learning algorithms
assisting the segmentation process, the need for
human participation is a vital thing in the selection
of features. The medical images have complex
information including uneven gray distribution,
large noise and easy deformation of the organ tissue,
which increase the difficulty of feature selection and
image segmentation.
Convolutional Neural Network, as one of the
most important models of deep learning, has
produced amazing segmentation results in medical
image segmentation.
The traditional CNN is two-dimensional
network. If it is directly extended to three-
dimensional, more parameters and calculating
processes are needed. The three-dimensional
network is more complicated, which requires longer
training time and more training data. The simple use
of two-dimensional data does not make use of three-
dimensional features, may result in decreased
accuracy. Adhish has adopted a compromise
solution for this purpose: use three 2D CNNs, which
are shown in the figure 5.
Figure 5 : An example for three-dimensional CNN [18]
Three 2D CNNs are responsible for the
processing of the xy, yz, and xz plane, and their
outputs are connected together through a Softmax
layer to produce the final output.
4.2 Discussion
The application of CNN is not limited to the above
cases, and it also includes object detection of self-
ICECTT 2018 - 3rd International Conference on Electromechanical Control Technology and Transportation
438

driving, handwritten character recognition, etc.
CNN’s vision is also very abundant, for example,
we can monitor the patients’ body image to analyze
the patients’ conditions through the camera instead
of wearing a certain sensor.
From a variety of applications and ideas, CNN’s
applications also need more in-depth studies.
5 CONCLUSION
In this paper, the Convolution Neural Network’s
history and structure are summarized. And then
several areas of Convolutional Neural Network
applications are enumerated. At last, some new
insights for the future research of CNN are
presented.
According to the above, on one hand, how to
improve the performance of the network deserves
more in-depth studying. On the other hand, CNN
should not be confined to the traditional field of
computer vision.
So, how to develop Convolution Neural Network
into the complicated, intelligent and real-time
application such as self-driving and traffic statistics
is also a problem worth noting.
REFERENCES
Krizhevsky A, Sutskever I, Hinton G E. ImageNet
classification with deep convolutional neural networks
[J]. Communications of the Acm, 2012, 60(2):2012.
Lawrence S, Giles C L, Tsoi A C, et al. Face recognition:
a convolutional neural-network approach [J]. IEEE
Transactions on Neural Networks, 1997, 8(1):98.
Kim Y. Convolutional Neural Networks for Sentence
Classification [J]. Eprint Arxiv, 2014.
Kandel E R. An introduction to the work of David Hubel
and Torsten Wiesel [J]. Journal of Physiology, 2009,
587(12):2733.
Yandong L I, Hao Z, Lei H. Survey of convolutional
neural network [J]. Journal of Computer Applications,
2016.
Fukushima K. Neocognitron: A hierarchical neural
network capable of visual pattern recognition [J].
Neural Networks, 1988, 1(2):119-130.
Lecun Y. LeNet-5, convolutional neural networks [J].
Ballester P, Araujo R M. On the performance of
GoogLeNet and AlexNet applied to sketches[C]//
Thirtieth AAAI Conference on Artificial Intelligence.
AAAI Press, 2016:1124-1128.
http://subscribe.mail.10086.cn/subscribe/readAll.do?colu
mnId=563&itemId=4111011
https://www.analyticsvidhya.com/blog/2017/06/architectur
e-of-convolutional-neural-networks-simplified-
demystified/
https://ujjwalkarn.me/2016/08/11/intuitive-explanation-
convnets/
ZHANG Yajun, GAO Chenqiang, LI Pei, LIU Jiang,
CHENG Hua. Pedestrian counting based on
convolutional neural network [J]. Journal of
Chongqing University of Posts and
Telecommunications:Natural Science Edition,2017,
29(2): 265-271.
Li X, Wang L, Sung E. AdaBoost with SVM-based
component classifiers [J]. Engineering Applications of
Artificial Intelligence, 2008, 21(5):785-795.
Prasoon A, Petersen K, Igel C, et al. Deep feature learning
for knee cartilage segmentation using a triplanar
convolutional neural network[C]// Medical Image
Computing & Computer-assisted Intervention: Miccai
International Conference on Medical Image
Computing & Computer-assisted Intervention. Med
Image Comput Comput Assist Interv, 2013:246.
http:// blog.csdn.net/taigw/article/details/50534376
Convolutional Neural Networks
439