Journal of Physics: Conference Series

PAPER • OPEN ACCESS

Research and Prospect of Image Recognition Based on Convolutional

Neural Network
To cite this article: Hanqing Hu et al 2020 J. Phys.: Conf. Ser. 1574 012161

View the article online for updates and enhancements.

This content was downloaded from IP address 45.82.99.47 on 04/07/2020 at 13:42

ICCASIT 2020 IOP Publishing
Journal of Physics: Conference Series 1574 (2020) 012161 doi:10.1088/1742-6596/1574/1/012161

Research and Prospect of Image Recognition Based on

Convolutional Neural Network

Hanqing Hu1,*, Jin Lyu2 and Xiaolin Yin3

1Beijing Information Science and Technology University, Haidian 100192, China
2Beijing Guozhixue Culture Co., Ltd, Chaoyang 100101, China
3BeiJing GuoXue Times Culture Co., Ltd, Fengtai 100070, China

*Corresponding author: [email protected]

Abstract. This paper compares common image recognition technology, introduces

three image recognition and classification technologies, and explains the most popular
deep learning image recognition algorithm based on CNN in detail. As a new method
of ANN, CNN combines ANN and deep learning technology. In the area of image
recognition, the advantage of CNN is particularly prominent. Compared with the
traditional image processing algorithm, CNN has higher recognition accuracy and
speed, and can directly process the original image to avoid the complicated
pre-processing image data. CNN is the first choice for image recognition.

Keywords: Image Recognition, Histogram of Direction Gradient, Principal

Component Analysis, Convolutional Neural Network

1 Introduction
Image processing is to transform the image into a digital matrix and store it in a computer, and process
it with a certain algorithm. The basis of image processing is mathematics, and the main task is the
design and implementation of various algorithms. At present, image processing technology has been
widely used in biomedicine, communication technology, remote sensing technology, cultural
creativity, industrial design and production and many other fields. Image recognition and classification
technology is to extract potentially hidden, useful and even unknown knowledge from a large number
of images. Image recognition and classification technology is the intersection of data mining and
analysis, machine/deep learning, image retrieval and image processing, machine/computer vision,
artificial intelligence technology, database/data warehouse and other disciplines [1].

2 Convolutional Neural Network

The purpose of deep learning is to establish a neural network to simulate the work of human brain
neurons. Therefore, deep learning theory is closely related to the development of neuroscience. David
Hubel and Torsten Wiesel have discovered that the human visual system is hierarchical. Specifically,
when a person sees an object, the picture first maps to the retina, then the edge features of the object
are extracted from the V1 area, the local features of the object are extracted from the V2 area, the V4
area is responsible for the overall features of the object, and the higher PFC (prefrontal cortex) is used

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
ICCASIT 2020 IOP Publishing
Journal of Physics: Conference Series 1574 (2020) 012161 doi:10.1088/1742-6596/1574/1/012161

for classification. It can be seen from the whole process of information processing in the brain that
information features are extracted from the bottom to the top, and the higher up, the more abstract the
features are. The whole process is as follows:

Fig. 1 Brain information extraction process

CNN has a structure of convolution, it can effectively reduce the computer memory consumption
and the number of network factors of deep-seated network, and alleviate the problem of model over
fitting. CNN is impacted by there key actions: local receptive field, share of weight and layer of
pooling.
CNN is a multilayer perceptron [2], it is successful because it adopts the way of local connection
and weight sharing. In this way, it not only reduces the weight number, makes the network
optimization easy, but also reduces the complexity of the model, that is, the risk of over fitting is
reduced. The advantage performance in the input of the network is the image more obvious, and
makes the image can be directly as network input, to avoid the traditional recognition algorithm of the
feature extraction and data reconstruction of the complex process, in the process of image processing
has a lot of advantages, such as network to extract image feature including color, shape，texture and
topology structure of the image and the processing of two-dimensional image, especially the
identification of displacement, zoom and other forms of distortion invariance applications has good
robustness and operation efficiency, etc. [3].
There are lots of factors in neural network, which are prone to over fitting and long training time.
However, compared with boosting, logistic regression, SVM and other methods based on statistical
learning theory (also can be seen as having a layer of hidden nodes or not the learning model with
hidden nodes, known as shallow model), has great advantages.
Deep learning’s two main viewpoints[4] are as follows: (1) The multi hidden layer ANN(artificial
neural network) has a good ability of the feature learning. The learned data could better react the
vitally important characteristics of the data, and make the data visualization or classification better; (2)
The difficulty of DNN(deep neural network) in training can be effectively overcome through
unsupervised training layer by layer .
Figure 2 shows the changes of the model. With the deepening of the model, the error rate of top-5
also becomes lower and lower, which has been reduced to about 3.5% at present. In the same
ImageNet data set, the error rate of human eyes is about 5.1%, which means that the recognition
ability of deep learning has surpassed that of human beings [5].

Fig. 2 The top-5 error rate of ILSVRC over the years

3 Algorithm

2
ICCASIT 2020 IOP Publishing
Journal of Physics: Conference Series 1574 (2020) 012161 doi:10.1088/1742-6596/1574/1/012161

3.1 Model Introduction

Fig. 3 CNN model introduction

Figure. 3 is a simple CNN structure. The orininal image of the 1st layer is convoluted to obtain the
characteristic image of the second layer with a depth of 3. The 2nd level feature map is combined to get
the 3rd level feature map with depth of 3. Repeat the above operations to get the characteristic map of
the 4th layer with a depth of 5. Finally, five feature maps, namely five matrices, are expanded by lines
and connected into vectors. The whole connection layer is BP neural network. Each characteristic
graph in the graph can look upon as a neuron arranged in matrix form, analogue to the neuron in BP
neural network.

3.2 Algorithm Implementation

(1) Convolution
After inputting a picture, it is necessary to convert it into a matrix, and the values of each element
of the matrix correspond to the values of each pixel of the picture. If there is a 5 × 5 image as fig4, a 3
× 3 convolution kernel as fig4 is used as convolution, and a 3 × 3 characteristic image could be got.
The convolution kernel is also called filter.
The specific operation process is shown in the figure 4 below:

Fig. 4 Convolution operation Fig. 5 Convolution complement zero

Yellow areas represent convolution kernels sliding in the input matrix, each slide to a location, the
corresponding digital multiplication and summation to get a element of matrix. Note that dynamic
convolution kernels in figure sliding a unit at a time, actually slip amplitude can be adjusted according
to need. If the sliding step of the convolution core is greater than 1, it is possible that the convolution
core cannot slide the edge of the picture exactly. To solve this problem, we can zero the matrix’s
outermost layer and fill the matrix with zero as shown in Fig.5.
The number of zeros can be set as needed. Zero Padding is a super factor which can be set, but
according to the size of the convolution kernel, it needs to be adjusted, the step length and the size of
the input matrix, so that the convolution kernel can slide to the edge.
Under normal circumstances, the input image matrix, and the back of the convolution kernel,
feature matrix is square, set size of the input matrix for 𝑤 here, convolution kernels for 𝑘 size, stride
for 𝑠, layer number of zero padding for 𝑝, characteristics of figure size after the convolution
calculation formula is:
𝑤 + 2𝑝 ― 𝑘
𝑤， = 𝑠 +1 (1)
(2) Pooling

3
ICCASIT 2020 IOP Publishing
Journal of Physics: Conference Series 1574 (2020) 012161 doi:10.1088/1742-6596/1574/1/012161

Pooling is also called Dwon sampling, as opposed to Up sampling. A pooling layer is generally
required to reduce the amount of data for the feature graph obtained by convolution. The pooling
operation is shown below:

Fig. 6 Pooling operation

Like convolution, Pooling also has a sliding kern, which can be called a sliding window. In figure
6, the sliding size window is 2×2 and the step length is 2. For each sliding area, the maximum value is
taken as the output. Such operation is called Max Pooling. Mean Pooling can also be used to output
the Mean value.
(3) The Whole Connection
The neurons of output layer and each neuron in the input layer are connected by each other. What
is the purpose of a full connection? Because the output of traditional network is classification, that is,
the probability of several categories or even a number -- category number, the full connection layer is
a feature of high purification, which is convenient for the final classifier or regression.
However, there are too many parameters for full connection. The current trend is to avoid full
connection as much as possible. One of the mainly methods is global average.
(4) Key Formulas
Input:
V = conv2(W，X，"𝑣𝑎𝑙𝑖𝑑") +b (2)
Output:
Y = φ(V) (3)
The above formula of input and output is for each of the convolution layer, each layer convolution
𝑊 has a different weight matrix, and 𝑊 𝑋, 𝑌 is matrix form. For the last full connection layer, set as
the first 𝐿 layer, the output is a vectorial 𝑦𝐿, in the form of expected output is 𝑑, has the total error
formula.
Total error:
1 2
E = 2‖𝑑 ― 𝑦𝐿‖2 (4)
Conv2() is the function of convolution operation in Matlab. The third parameter valid indicates the
type of convolution operation. The convolution method introduced above is the valid type. 𝑊
convolution kernel matrix, 𝑋 as input matrix, 𝑏 is biased, 𝜑 (𝑥) is the activation function. Total error of
𝑑, 𝑦 were expected output and network output vector.||x||2 represents the 2-norm of the vector 𝑥, the
calculation expression is:
1

‖𝑥‖2 = (∑𝑥2𝑖 )2 (5)

The input and output formulae of the all-connected layer neurons are exactly the same as the BP
network. 𝜑 is a kind of activation function[5].

3.3 Local Connection Properties of CNN

The convolutional neural network belongs to the local connection network, which is based on the deep
study of natural images. Natural images have the property of local region stability, and the statistical
characteristics of a local region are similar to those of other adjacent local regions. Therefore, the

4
ICCASIT 2020 IOP Publishing
Journal of Physics: Conference Series 1574 (2020) 012161 doi:10.1088/1742-6596/1574/1/012161

characteristics of a local region learned by neural network from natural images are also suitable for
other adjacent local regions of images.
Compared with local connected network and full connected network, weight sharing network has
great advantages. The weight sharing network can reduce the number of training parameters, so as to
simplify the network structure and have wider application space. If the input image of 1000×1000 is
taken as an example, and the number of nodes in the convolutional layer is 106, 1012 weights are
required for the full connection. Assuming that the fully connected network adopts a local receptive
field of 10×10, the weights need to be reduced to 108. If the system has 100 filters, the weight sharing
network can reduce the number of weights to 104 based on the local receptive field. Thus, the weight
sharing network and the local connection network can greatly reduce the network parameters, simplify
the network structure, and improve the efficiency of image recognition.

4 Application Scenarios
CNN are widely used in the field of images, are currently less interpretable, just like a "black box".
Although they are constantly evolving, it is hard to say why the model performs so well.
Convolutional neural network has got big achievements in the field of image processing and
recognition. On the international standard ImageNet dataset, lots of successful models are based on
convolutional neural network. One of the advantages of the CNN technology algorithm is that it can
directly analyze the original image and avoids the original image’s complicated pre-processing[6].
Every major breakthrough in image recognition has involved the use of CNN as well as many derived
with this network model, can be directly to the image data as input, not only without manual for image
preprocessing and extra complex operations, such as feature extraction, and with its unique way of
fine-grained feature extraction, makes the image processing to reach the level of human.
With the further development of random technology, CNN has been a large number of applications
applied in face recognition, education image processing, intelligent driving, intelligent security, text
recognition, human-computer interaction, image search, and intelligent home[7][8].
The strength of CNN is that it can map low-dimensional low-level features to high-dimensional
high-level features. Therefore, all data that satisfy local correlation can be theoretically processed by
CNN, such as voice and text. In terms of natural language processing, CNN can be used to do some
basic tasks, such as part-of-speech tagging, entity recognition, text classification, etc., as well as some
cutting-edge tasks, such as machine translation, chatbot, etc.

5 Summary and Outlook

In recent years, the image recognition technology is increasingly hot, every year at the rocket speed to
update the new technology and achievements. CNN has turned into the first solution for image
classification. Its image recognition accuracy is so high that it can be used in a wide range of
applications across different platforms, such as smartphones, security systems and driver assistance
systems.
With the improvement of CNN depth and network structure, the recognition accuracy and speed of
CNN in image recognition have been improved, and the field of image recognition has been gradually
expanded with increasingly powerful functions. It is very difficult to determine which network
structure to use, how many layers to use, how many neurons to use is suitable. Detailed knowledge is
still needed to select reasonable values, such as learning rate, regularization intensity, etc., and the
application cost is high. Moreover, due to the lack of universality of network structure, there are great
limitations in solving problems. It is found that the CNN has a very broad application prospect in the
area of image recognition. Therefore, it is necessary to research all the image recognition system，to
optimize the network structure and the depth of CNN.
(1) In the process of image recognition, whether the filter size is appropriate or not has a direct
impact on the training process and recognition accuracy. Therefore, in order to make the image
recognition result better, it is necessary to select the filter with the most appropriate size.

5
ICCASIT 2020 IOP Publishing
Journal of Physics: Conference Series 1574 (2020) 012161 doi:10.1088/1742-6596/1574/1/012161

(2) When using CNN for image recognition, the network depth of different problem selections is
often different, and the depth value needs to be determined by manual pre-selection and
experimentation, which limits the universality of the network structure. Therefore, in practical
applications, we need to select an approximate general network structure depth for specific image
recognition problems.
(3) The application of CNN in image recognition mainly achieves the application effect through
data set training, which has great limitations for different data sets. Therefore, the CNN needs to train
the existing data set according to different problem sets in order to get the same result. If the
distribution of the training data set is different from that of the test data set, it is difficult for the
convolutional neural network to obtain a good recognition result.
In the image processing based on CNN, a complete normal theory has not been formed. At present,
many identification systems design the depth and level of the network based on a specific database,
and find the best parameters and optimization algorithm through continuous exploration. Human
reasons are relatively prominent, and there is no systematic theoretical explanation of factors affecting
the recognition effect of the CNN. In particular, when classifying and recognizing natural images, the
selection of the initial state parameters of the CNN and the optimization algorithm will have a big
effect on the network training [9]. If your validation set is more accurate than your training set, the
model is under-fitted. Overfitting occurs when your model overfits the training set. Selection will
result in the network not working, or it may over-fitting, under-fitting and any other problems.
We need to understand deeply CNN meanings and roles of each part, adjust various parameters to
optimize and deepen the network, make the network to obtain more information, or even to add our
innovation in the structure to solve various problem.

Acknowledgments
This research was supported by National Key R&D Program of China (Grant No.2017YFB1400400).

References
[1] Meng-xue Xu: An overview of image recognition technology based on deep learning. Computer
Products and Circulation 1, 213-213(2019).
[2] JING L, CHENG J H, SHI J Y, et al：Improvement. 2012.
[3] Liang Chang et al.: Convolutional neural networks in image comprehension. Automation 9,
1300-1312(2016).
[4] HINTON G E et al.: A fast learning algorithm for deep belief nets. Neural Computation 7,
1527-1554(2014).
[5] Xian-chang Chen: Deep learning algorithm and application research based on convolutional
neural network. Zhejiang Gongshang University,2014.
[6] Lin Zhang et al.: A review of research on convolutional neural networks. Chinese Journal of
Computers 40, 1229-1251(2017).
[7] Quan-Sen Sun et al.: A new method of feature fusion and its application in image recognition.
Pattern Recognition 12, 2437-2448(2005).
[8] Misgana Negassi et al.: Application of artificial neural networks for automated analysis of
cystoscopic images. World Journal of Urology 1, 2020.
[9] XU K, FENG D, MI H, et al.: Mixup-Based Acoustic Scene Classification Using Multi-channel.
Convolutional Neural Network, 2018.