2021 IEEE Asia-Pacific Conference on Image Processing, Electronics and Computers (IPEC)

Research on Application of Deep Learning

Algorithm in Image Classification
2021 IEEE Asia-Pacific Conference on Image Processing, Electronics and Computers (IPEC) | 978-1-7281-9018-1/20/$31.00 ©2021 IEEE | DOI: 10.1109/IPEC51340.2021.9421185

Zhiping Wang1, *, Cundong Tang1, 2, Xiuxiu Sima1, Lingxiao Zhang1

1
Software Engineering College, Nanyang Institute of Technology, Nanyang 473004, China
2
School of Information Science and Technology, Northwest University, Xi’an 710127, China
e-mail: [email protected]

Abstract—Image classification is an important research prospect, and strengthening the research on image
direction in the field of computer vision. Image classification recognition technology plays a vital role in the development
algorithm can distinguish different kinds of pictures by of artificial intelligence and computer vision [6]. In essence,
classifying the features extracted from the original pictures. In deep learning is actually a technology to simulate and
essence, deep learning is actually the technology of simulating analyze the human brain through the construction of deep
and analyzing the human brain through the construction of neural network, or to learn and interpret the relevant data by
deep neural network, or learning and interpreting related data simulating the human brain [7]. Image recognition is a
by simulating the human brain. As an important information
technology to recognize objects and targets of various
carrier, image has become an important form for people to
patterns after analyzing, understanding and processing
acquire and transmit information. For massive images, the
number of images that each user really needs is very small, so
related images [8]. Traditional image classification
it is of great significance to find the images that users need algorithms, such as image classification based on manual
within the effective time. Based on this, this paper analyzes the marking and key point description, are not only
research and application of deep learning in image recognition, cumbersome and time-consuming, but also greatly affected
hoping to further enhance the application effect of deep by human subjective factors, and the classification accuracy
learning and make it play a greater role in the field of image is not high. The image classification algorithm based on
recognition. multi feature fusion and deep learning can get more efficient
image classification effect. As an important information
Keywords—Image classification, Computer vision, Deep carrier, image has become an important form for people to
learning obtain and transmit information. For massive images, the
number of images that each user really needs is very small,
I. INTRODUCTION so it is very important to find the images that users need in
With the increasing types of image acquisition the effective time [9]. For complex image semantic features,
equipment and the continuous maturity of image processing efficient deep learning algorithm has important research
technology, a large number of images are generated in every significance for fast and effective extraction and analysis of
field every day, which makes the images exist in a massive image semantic information [10]. This paper analyzes the
form [1]. Image classification is an important research research and application of deep learning in image
direction in the field of computer vision. Image recognition, hoping to further improve the application effect
classification algorithm classifies the features extracted from of deep learning, so that it can play a greater role in the field
the original image to achieve the purpose of distinguishing of image recognition.
different categories of images. Artificial intelligence is
widely used in various fields of society, which not only II. DEEP LEARNING IN IMAGE RECOGNITION
improves the development level of industry modernization Deep learning originates from artificial neural network,
and intelligence, but also brings great convenience to which is a network structure with multiple hidden layers. It
people's life [2]. With the development of Internet usually uses CNN (Convolutional Neural Network Model)
technology, a large number of unlabeled image data flow to classify static images. With the further development of
onto the network, and deep learning algorithm can extract social science and technology, scientists' research on
abstract feature representation from these unlabeled data by artificial neural network has gradually entered a bottleneck
using multi-layer nonlinear transformation for image period, and back propagation algorithm has been born.
classification [3]. As a key component of machine learning, Through this technology, artificial neural network can adapt
deep learning plays an important role in the development of to more complex data calculation, which makes the learning
modern artificial intelligence [4]. The application of deep level of artificial neural network gradually deepen. Neurons
learning greatly improves the accuracy of image in the same feature plane can share weights, and the shared
classification, and its performance even surpasses that of weights are convolution kernels. Convolution kernel can
human beings in some fields [5]. The improvement of the obtain reasonable weights through learning during network
accuracy is based on the huge amount of calculation. With training, and at the same time, sharing weights can reduce
the increase of the number of layers and the number of the connections between networks at all levels, thus
nodes in each layer, the amount of calculation required by effectively reducing the fitting risk [11]. In the task of image
the training network also increases sharply. classification, database plays an important role, especially in
today's era dominated by deep learning. The quality of
In the field of image recognition, deep learning is a very
database often directly determines the quality of the trained
important technical means, which has a broad application

978-1-7281-9018-1/21/$31.00 ©2021 IEEE 1122 Dalian, China

April 14-16, 2021

Authorized licensed use limited to: East Carolina University. Downloaded on June 19,2021 at 02:23:20 UTC from IEEE Xplore. Restrictions apply.
model to a great extent. Ideally, the pictures in the training for collaborative training. The image classification method
set and the test set should be adjusted to a unified posture. based on convolution neural network simulates the human
However, after carefully observing the database, we find brain vision system to classify the images through the model
that this is impossible, because the different shooting angles structure with convolution layer and downsampling layer
and different picture scales lead to excessive differences in alternating in turn. Convolution layer enhances the original
the postures of targets in many pictures. The actual driving signal and improves the signal-to-noise ratio to some extent.
process of the motion system is obtained through visual At the same time, according to the principle of local
behavior analysis, and a unified visual and behavioral model correlation of images, the downsampling layer samples
is formed, as shown in Figure 1. images from different fields, extracts useful information and
reduces the amount of data.
III. APPLICATION OF DEEP LEARNING IN IMAGE
CLASSIFICATION AND RECOGNITION
A. Image Recognition
Image classification involves many technologies, such as
image preprocessing, image feature extraction, image
category labeling, and image classifier construction, all of
which will have a certain impact on the image classification
effect. In the process of RBM training, with the increase of
training times, the network model gradually converges.
When the error increment is getting smaller, it shows that
the current learning rate is more appropriate. If the learning
Figure 1 Vision and behavior model
rate is blindly increased, the error may increase again, thus
causing the network to oscillate. If the deep learning
If we want to improve the performance of deep technology can be applied in remote sensing image
convolutional neural networks, the most intuitive method is classification, the classification effect can be effectively
to increase the depth and width of deep convolutional neural improved by constructing the deep learning model
networks, where depth is the number of layers and width is reasonably and applying the corresponding optimization
the number of nodes in each layer. Deep learning algorithm. In the working process of image classification
emphasizes the depth of network, aims at learning features method based on deep learning algorithm, there are many
at different levels, and then highlights the importance of key technologies, among which features and classifiers are
learning, that is, learning high-level features from low-level the most critical [14]. For large-scale images, the
features. In order to solve complex function problems by classification effect of support vector machine is low, and it
solving relatively few parameters, a multi-layer network is seems a little powerless to classify images accurately. Deep
constructed, so that the input data takes the sequential action learning algorithm is a new type of machine learning
of multiple hidden layers, that is, the output of the previous algorithm developed on the basis of traditional neural
hidden layer as the input of the next hidden layer, and the network, which can learn large-scale data through multi-
feature maps of different levels are obtained [12]. The level network, accelerate the modeling speed, and be widely
cooperative training algorithm itself is designed for multi- used in speech recognition, biological information and other
view data, but later some improved algorithms appeared fields. In the process of weight updating, inertia factor is
when dealing with single-view data, some of which use introduced, which makes RBM model have strong global
different classification learning algorithms, some use search ability with larger inertia weight in the early stage
different data sampling methods, and even use different and good local mining ability with smaller weight in the
parameter settings to produce significant differences. later stage. Figure 2 shows the structure of digital image
analysis system.
The images in the database obtained translation
invariance and rotation invariance through the above The attenuation rate of learning rate is adaptively
positioning and rotation operations. According to the adjusted by the error increment in each iteration, and the
traditional image processing theory, these images still lack error control factor is introduced to control the influence of
scale invariance. Each node in the hidden layer is related to error increment on learning rate. At the same time, inertia
a certain extent, and the input in the hidden layer involves factor is set. In the early stage of the algorithm, large inertia
the output of the previous period on the hidden layer and the weight is used to have strong global search ability, and in
output of the input layer. The generation countermeasure the later stage, small weight is used to have good local
network in the form of deep convolution and the generation mining ability, thus improving the convergence speed and
countermeasure network in the form of Laplace pyramid stability of RBM training. In the practical application
realize the combination of unsupervised countermeasure process, the stacked automatic encoder also operates layer
network and supervised learning convolution neural network, by layer, and each layer has different data processing and
thus determining a set of structures that can realize stable expression methods, and the data characteristics caused by
training [13]. Bifurcation-based collaborative training these differences will be summarized and analyzed, thus
method is a commonly used semi-supervised learning realizing data preprocessing. At the present stage, the
method, which can prove that if the view has sufficient importance of ImageNet classification for deep learning is
conditional independence, the original weak classifier can also increasing, and at the same time, a network structure in
be upgraded to a strong classifier with higher generalization the form of AlexNet has been formed. Compared with the
performance by adding a large number of unlabeled samples traditional convolution network, this network structure has
many advantages. AlexNet can effectively reduce the

1123

Authorized licensed use limited to: East Carolina University. Downloaded on June 19,2021 at 02:23:20 UTC from IEEE Xplore. Restrictions apply.
computational complexity, and can effectively obtain the sparse neuron output.

Figure 2 Digital image processing and analysis system structure

B. Image Classification The accuracy of image classification depends on feature

Global features and local features are the main forms of extraction, but the ability of the underlying features to
features in image classification. The global features of an describe images is weak and the classification effect is poor.
image are structural features, color features and texture Therefore, the image is described by multi-feature fusion,
features. With the rapid development of modern society, the and the complementary information and redundant
application scope of image recognition technology has been information are eliminated, thus improving the image
further expanded, and it has been widely used in the field of classification effect. As shown in Figure 4, the image path is
transportation. This kind of traffic image recognition determined by motion detection.
technology is used in traffic sign recognition, lane departure
warning and license plate recognition, and it also brings a
lot of convenience for people's transportation. Color feature
is characterized by the color of the image surface. It is a
global feature to provide the basis for image classification
by histogram method, but it is difficult for image feature to
obtain satisfactory local information, so it needs to be
combined with other features. Deep learning has also been
applied properly in video image analysis, but it is still in its
infancy, so it is easy to describe video static images by deep
learning, and the related deep learning model can be
obtained from ImageNet [15].
The study of extreme speed learning machine with class
constraints is a study of distributed expression features. In
this case, the data can be mapped to different spaces, and
then through feature combination, a feature expression with
discriminatory information is formed, that is, a distributed
expression. Figure 3 shows the frame of sports detection and
tracking system using image error technology.

Figure 4 Motion detection to determine the image path

The deep learning algorithm has the highest accuracy in

image classification, and the image classification results are
very stable, mainly because the generalized regression
neural network algorithm can accurately describe the
mapping relationship between image classification input
feature vectors and image categories, and can distinguish
various categories of image classification, thus obtaining the
optimal image classification results. In the process of
moving image recognition, using deep learning to build a
model of moving image recognition can not only improve
the efficiency of moving image recognition, but also
improve the technical level of sports training. At present, the
Figure 3 The framework of sports detection and tracking system using application of deep learning in the field of image
image error technology classification and recognition is relatively mature. Therefore,

1124

Authorized licensed use limited to: East Carolina University. Downloaded on June 19,2021 at 02:23:20 UTC from IEEE Xplore. Restrictions apply.
rock features can be extracted by building a model, and then [2] Tang Sipeng, Chi Hongyuan, Zhang Peiwei, etc. The design and
intelligent classification can be realized, so that the image application of deep learning to identify optical network unit failures.
Computer Technology and Development, vol. 277, no. 5, pp. 217-221,
results are more accurate. Studying the application of deep 2020.
learning in image recognition can further enhance the [3] Liu Hongchao, Dong Anguo. Hyperspectral remote sensing image
research depth of related fields, which has a very positive classification algorithm based on non-local feature fusion. Progress in
effect on the effective application of various technologies Laser and Optoelectronics, vol.665 no. 6, pp. 195-201, 2020.
and the development of related fields. [4] Zeng Rui, Chen Duisheng. Hyperspectral remote sensing image
classification combined with dual deep learning features. Small
IV. CONCLUSIONS microcomputer system, vol. 39, no. 2, pp. 396-400, 2018.
[5] Zhou Zhongyi, Wu Jin, Zhu Lei. Dew and frost image classification
Studying the application of deep learning in image based on multi-path feature fusion and deep learning. Computer
recognition can further enhance the research depth of related Applications and Software, vol. 035, no. 10, pp. 205-210, 2018.
fields, and has a very positive effect on the effective [6] Lin Sisi, Ye Dongyi, Chen Zhaojiong. Flower image deep learning
application of various technologies and the development of classification algorithm based on multi-feature fusion. Small
related fields. This paper mainly summarizes deep learning, microcomputer system, vol. 39, no. 7, pp. 1446-1450, 2018.
analyzes the challenges faced by this technology in the [7] Hongtao Shang, Guoliang Shi. Hyperspectral remote sensing image
application of image classification, and puts forward some classification method based on SSAE deep learning feature
practical application directions. The images in the database representation. Computer Measurement and Control, vol. 26, no. 10,
pp. 297-300, 2018.
obtained translation invariance and rotation invariance
[8] Li Qiong, Bai Zhengyao, Liu Yingfang. Deep learning classification
through the above positioning and rotation operations. method for diabetic retina images. Chinese Journal of Image and
According to the traditional image processing theory, these Graphics, vol. 23, no. 10, pp. 166-175, 2018.
images still lack scale invariance. In order to solve some [9] Kang Xiaodong, Wang Hao, Guo Jun, et al. Unsupervised deep
problems existing in the current image classification process, learning color image recognition method. Computer Applications, vol.
such as high misclassification probability, an image 35, no. 9, pp. 2636-2636, 2015.
classification method based on deep learning algorithm is [10] Xu Feng, Wang Haipeng, Jin Yaqiu. Application of Deep Learning in
proposed, and an image classifier is established by SAR Target Recognition and Ground Object Classification. Journal of
Radars, vol. 6, no. 2, pp. 136-148, 2017.
introducing generalized regression neural network of deep
learning algorithm. The accuracy of image classification [11] Lin Zewen. Deep learning and its application in image classification
and recognition. Information and Computer (Theoretical Edition), vol.
depends on feature extraction, but the ability of the 406, no. 12, pp. 151-152, 2018.
underlying features to describe images is weak and the [12] Zhang Ye, Li Mingchao, Han Shuai. Automatic lithology recognition
classification effect is poor. Therefore, the image is and classification method based on deep learning of rock images.
described by multi-feature fusion, and the complementary Acta Petrologica Sinica, vol. 34, no. 2, pp. 81-90, 2018.
information and redundant information are eliminated, thus [13] He Xueying, Han Zhongyi, Wei Benzheng. Automatic classification
improving the image classification effect. Studying the of breast cancer pathological images based on deep learning.
application of deep learning in image recognition can further Computer Engineering and Applications, vol. 54, no. 12, pp. 121-125,
2018.
enhance the research depth of related fields, which has a
[14] Zheng Yuanpan, Li Guangyang, Li Ye. A review of the application of
very positive effect on the effective application of various deep learning in image recognition. Computer Engineering and
technologies and the development of related fields. Applications, vol. 55, no. 12, pp. 20-36, 2019.
[15] Peng Bo, Cai Xiaoyu, Tang Ju, et al. High-altitude video vehicle
ACKNOWLEDGEMENTS recognition based on morphological detection and deep learning.
This work was supported in part by the Industrial Transportation System Engineering and Information, vol. 19, no. 6,
pp. 45-51, 2019.
Science and Technology Research Project of Nanyang under
Grants KJGG002, and KJGG011, andKJGG2018135, and
2018-ZDJH-219.
REFERENCES
[1] Ge Mengying, Yu Chongzhong, Zhou Lan, et al. Deep learning image
classification algorithm based on collaborative semi-supervised.
Computer simulation, vol. 36, no. 2, pp. 206-210, 2019.

1125

Authorized licensed use limited to: East Carolina University. Downloaded on June 19,2021 at 02:23:20 UTC from IEEE Xplore. Restrictions apply.