This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3021590, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.Doi Number

Image Recognition Technology Based on

Machine Learning
Lijuan Liu1,a, Yanping Wang2,b and Wanle Chi3,4,5*
1College of Computer Science, Neijiang Normal University, Neijiang 641112, Sichuan, China
2INFORMATION CENTER, China Tobacco, Zhejiang Industry Co. Ltd, Hangzhou 310008, Zhejiang, China
3Department of Information Technology, Wenzhou Polytechnic, Wenzhou 325035, China

4Faculty of Information and Communication Technology, Technical University of Malaysia Malacca, Malacca 76100, Malaysia

5Wenzhou University, Wenzhou 325000, China

[email protected]

[email protected]

*Corresponding author: Wanle Chi (e-mail: 00884070 @wzu.edu.cn)

This work was supported by Key scientific research project of Neijiang Normal University (NO.18ZA06); Natural Science Foundation of Zhejiang
Province.

ABSTRACT With the development of machine learning for decades, there are still many problems
unsolved, such as image recognition and location detection, image classification, image generation, speech
recognition, natural language processing and so on. In the field of deep learning research, the research on
image classification has always been the most basic, traditional and urgent research direction. At the same
time, computer intelligent image recognition technology is also conducive to gradually better respond to the
development of international indicators, and promote the development and progress of various fields.
Therefore, image processing technology based on machine learning has been widely used in feature image,
classification, segmentation and recognition, and is a hot spot in various fields. However, due to the
complexity of video images and the distribution of objects in different application backgrounds, the
classification accuracy becomes important and difficult. In the paper transportation industry, image
recognition technology is applied to license plate recognition to extract license plate from complex
background, segment license plate characters and recognize characters, and construct a machine learning
non license plate automatic generation algorithm, which may improve the efficiency of non license plate
recognition. The diversity and high generation speed of license plate training sample set can achieve the
purpose of effectively training strong classifier. By using genetic algorithm to optimize BP neural network
to classify license plate information, the anti-interference ability and license plate recognition accuracy are
improved to a certain extent.

INDEX TERMS Artificial Intelligence, Image Preprocessing, Image Recognition, Machine Learning

I. INTRODUCTION Therefore, how to improve the classification method to

Machine Learning (ML, Machine Learning) [1-5] is a improve the classification accuracy and classification effect
fundamental and critical issue in the field of image of the image of the ground object is a very meaningful and
processing [6, 7], especially in the field of massive image difficult research topic. With the development of machine
processing, machine learning methods can be from complex learning and the introduction and improvement of various
data [8]. The main features of the image are separated [9], machine learning algorithms, machine learning is of great
so that image recognition can be reasonably applied in significance to various application fields in human life.
various industries and fields. Image processing technology Especially with the rapid development of modern
based on machine learning has been widely used in image technology and the application of video images in various
classification, segmentation, and recognition [10]. It is a fields of life, machine learning is particularly important for
hotspot of research and research in various fields. However, the processing of video images. At present, various machine
due to the complexity of image distribution and different learning algorithms have been maturely applied to signal
application backgrounds, the improvement of image processing in engineering, but in video image processing,
classification [11-12] has become the focus and difficulty. there is still a broad application space. The application of

VOLUME XX, 2017 1

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3021590, IEEE Access
Research on Image Recognition Technology Based on Machine Learning

machine learning to target image classification technology recognition technology. This paper identifies the license plate
is related to the development of various industries in China. based on the machine learning method, and classifies the
Therefore, the application of machine learning in target sample using BP neural network trained by genetic algorithm
image classification has become a very important research [30]. Compared with the genetic algorithm under different
topic. fitness, the solution with higher accuracy is obtained.
Computer image recognition technology [13-17] is actually
the abbreviation of computer image processing and II. PROPOSED METHOD
recognition technology, also known as infrared technology.
The core of this technology is computers and information. A. MACHINE LEARNING
These two technologies are the most developed in the world. Machine Learning (ML) is a multidisciplinary subject
The former is the real carrier of technology. It undertakes the involving many disciplines such as probability theory,
analysis and processing of the image, and then carries on the statistics, approximation theory, convex analysis, and
different localization correctly. The object of the information. algorithm complexity theory. Specializing in how
Infrared technology can be said to be the product of social computers simulate or implement human learning behaviors
development and the progress of the times [18]. The image is to acquire new knowledge or skills and reorganize existing
input into the neural network, and the loss function is knowledge structures to continuously improve their
minimized by using the forward propagation and back- performance. It is the core of artificial intelligence, and it is
propagation error algorithms of deep learning. After the the fundamental way to make computers intelligent [31]. Its
weight is updated, a better recognition type is obtained. Then, application spans all fields of artificial intelligence. It
mainly uses induction, synthesis rather than deduction.
the trained model is used to predict the new image. The flow
Simply put, machine learning is a process of extracting
chart is shown in Figure 1.2. General pattern recognition
useful information from unordered data. It spans multiple
system includes three important parts: image preprocessing,
disciplines such as computer science, engineering, and
feature extraction and classifier. In traditional image
statistics and requires multidisciplinary knowledge. In the
recognition algorithm, they are separated from each other. In
Internet age, people create and collect a large amount of data.
the framework of convolutional neural network, convolution
How to extract valuable information from these data is a topic
is used to extract features directly, and then the classification
worth studying. Now is also the era of “data is king”,
results are fed back to the classifier, and the model is jointly
companies are crazy to collect user data, personal
optimized by batch gradient descent. The process of
information, usage habits, search records, watch records and
computer preprocessing [19-21] is mainly to separate the
even email content... hope to find user preferences and tap
image area and background area in the image to be
users’ needs . Who has the data, who has the next
recognized, refine the image, enhance the image binarization,
opportunity. However, it is not enough to have such data. The
and improve the speed and efficiency of computer intelligent
massive data has exceeded the feasibility of direct
image recognition post-processing. In order to restore the
calculation. To extract information efficiently from it, a
authenticity of the image and reduce the false features of the
special learning algorithm is needed. This is the role of
image as much as possible, the unique features of the image
machine learning. The “machine learning period” is also
can be expressed in numerical form. With the development
divided into three stages. In the 1980s, connectionism was
and progress of technology, digital image is gradually used in
more popular, representing work with Perceptron and Neural
the field of image recognition. The advantages of digital
Network. In the 1990s, statistical learning methods began to
processing technology provide the basis for the further
occupy the mainstream stage. The representative methods
development of image recognition. In these two development
were Support Vector Machine [32]. In the 21st century, deep
stages, infrared technology explored a series of successful
neural networks were proposed. Connectionism has never
methods through the research and application of artificial
been seen, with the increasing amount of data and computing
intelligence [22,23], and finally realized the effective
power. Many AI applications based on Deep Learning have
identification of information. Since then, this technology has
matured.
been widely used. Image recognition is widely used in traffic
Machine learning is a general term for a class of
field. In traffic construction, image recognition technology is
algorithms that attempt to mine the implicit rules from a large
mainly used in intelligent transportation system [24]. Vehicle
amount of historical data and use them for prediction or
information detection [25-28] has greatly promoted the
classification. More specifically, machine learning can be
development of transportation modernization.
seen as looking for a function, and input is sample data. The
Vehicle detection is an important part of the effective
output is the desired result, but this function is too
operation of the traffic monitoring system, but if you want to
complicated to be formally expressed. It is important to note
better identify and track the vehicles in the traffic network,
that the goal of machine learning is to make the learned
you need to correctly segment the vehicle and obtain the
functions work well for “new samples,” not just for training
target area [29]. The same is true for license plate
samples. The ability of the learned function to apply to a new
recognition. This method can be carried out well by image
sample is called generalization capability. In terms of scope,

2 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3021590, IEEE Access
Research on Image Recognition Technology Based on Machine Learning

machine learning is similar to pattern recognition, statistical of image processing technology is obtained. The whole
learning, and data mining [33]. At the same time, the principle is not complicated, that is, the view between people
combination of machine learning and processing techniques can be regarded as completing an image recognition
in other fields forms an interdisciplinary subject such as technology, and then the acquired information is analyzed in
computer vision, speech recognition, and natural language the brain according to the impression. The principle of
processing. Therefore, in general, data mining can be artificial intelligence image recognition technology is the
equivalent to machine learning. At the same time, what we same as that of computer processing data; therefore, simple
usually call machine learning applications should be image data information extraction can be performed by
universal, not only limited to structured data, but also to computer, but when the amount of information is large, the
applications such as images and audio. recognition rate of image recognition technology will
Machine learning is widely used in many fields. For decrease, and relevant personnel are analyzing The principle
example, speech recognition is a combination of audio of image recognition technology should look for more
processing technology and machine learning. Speech optimized methods for innovation, so as to improve the
recognition technology is generally not used alone, and quality and efficiency of image processing. The image
generally incorporates related techniques of natural language recognition technology in artificial intelligence has the
processing. The current related applications are Apple’s voice advantages of convenience and intelligence. The advantages
assistant siri and so on. In image processing techniques, of the technology directly determine the application quality
images are processed into inputs suitable for entry into a and effect of image recognition technology in the
machine learning model, and machine learning is responsible development of science and technology. First of all, from the
for identifying relevant patterns from the images. There are perspective of intelligence, the most obvious advantage of
many applications related to computer vision, such as Baidu artificial intelligence image recognition technology is
map, handwritten character recognition, license plate intelligence. Compared with traditional image processing
recognition and so on. This field is very promising and is also technology, it shows a clear difference. This function can
a hot research direction. With the development of deep realize intelligent selection and recognition when processing
learning in the new field of machine learning, the effect of pictures, such as the face unlocking function in the mobile
computer image recognition has been greatly promoted, so phone, which is very similar to the intelligent recognition
the future development of computer vision industry is function in image processing, that is, the face unlocking can
immeasurable [34]. be permanently used as long as the face unlocking is
completed. Intelligentization not only enables image
B. ARTIFICIAL INTELLIGENCE recognition and other functions, but also enables self-analysis
In the process of using computer vision algorithm to and preservation. Secondly, from the convenience of graphic
simulate human image recognition, researchers have recognition technology, with the application of image
proposed many different image recognition models. Among recognition technology, it has created a lot of excellent
them, the image recognition algorithm based on template services for people’s life and work. In this technology, people
matching is the most widely used. Whether the target image do not need to perform complex image processing to achieve
features in the image database are consistent with the target the purpose, such as brushing face punching, brushing face
features to be matched is determined by matching the target unlocking, etc., which bring convenience to people’s lives.
image with the predicted image. The principle of image With the development of society, image recognition
recognition technology in artificial intelligence is combined technology has become more and more popular, and it is
with the algorithm principle of computer processing data, so more convenient to use.
the simple image data information extraction and analysis can Because the image recognition technology is implemented
be combined with the computer, but in the case of fuzzy based on artificial intelligence, the image recognition process
image information or large amount of information in the of the computer is almost the same as the human brain image
image, the recognition efficiency is high, and the image recognition process. The biggest difference is that the
recognition technology may be reduced. Therefore, when computer image recognition is displayed in the form of
analyzing the principle of image recognition technology, we technology. The specific artificial intelligence image
also need to find a better and more convenient image recognition process is as follows. First, information data is
recognition technology. Its principle is to change the image obtained. Information collection is a prerequisite for image
recognition technology, make the principle of image recognition. It mainly converts various special signals into
recognition technology more simple, and achieve better in electrical signals through sensors, and then obtains the
function and image processing. required information and data from them [35]. However, the
The principle of image recognition technology in artificial information acquired in image recognition technology
intelligence is to use computer to process pictures, and then belongs to the special data of images. The data must be able
extract the information in pictures. Through the analysis and to distinguish the gaps between the graphics. Second,
experiment of Chinese professionals, the technical principle information data is preprocessed. This stage is mainly to

VOLUME XX, 2017 3

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3021590, IEEE Access
Research on Image Recognition Technology Based on Machine Learning

smooth, transform and other images, in order to highlight the effect and the judgment result to meet the specific application
important information of the image itself. Third, feature requirements. This processing technique can be called image
extraction and selection. This is the key content of image enhancement processing. technology. The image
recognition technology, especially in the recognition mode, enhancement technology can be roughly classified into two
the actual operation requirements are higher, which also categories, a spatial domain method and a frequency domain
directly determines whether the image can be successfully method, according to different positions of objects processed
recognized and whether the extracted features can be stored. by the enhancement technique. The spatial domain-based
Fourth, classifier design and classification decisions. This is algorithm refers to the gray value of the original pixel directly
the last step of image recognition. This part mainly processed when the image is based on the image’s own plane.
formulates the recognition rules according to the operation The frequency domain method is to enhance the image on
procedure, and recognizes the image according to the another transform domain of the image. Histogram
standard instead of the chaotic recognition. The purpose is to equalization is a processing method that enhances the
improve the recognition degree of the image processing, operation of digital images based on probability theory. The
thereby improving the efficiency of image evaluation. histogram, also known as the mass distribution map and
histogram, is a statistical graph based on the report. The
C. IMAGE PREPROCESSING histogram of a digital image is a distribution of the total
In the process of image acquisition, it is often subject to number of pixels of different gray values in an image.
various external conditions and random interference. Such Through the histogram of an image, we can see the brightness
directly acquired images often contain complex useless of the grayscale distribution of the pixel of this image. The
backgrounds or redundant data, which interferes with the grayscale value of the histogram of the over-dark image is
further application of images. Therefore, some necessary concentrated at the lower part, the overall over-bright image,
pre-processing techniques need to be performed on the its histogram The body of the graph is distributed at a higher
original data image. Commonly used image processing gray value. The method of histogram equalization is to
operations include color image grayscale technology, image transform the histogram of the original image by gradation
enhancement technology, image restoration technology, transformation and to correct the stretching according to a
image segmentation technology, smoothing and sharpness,
certain rule, and obtain a new histogram image with stable
and the like [36]. In order to facilitate computer processing,
gray value distribution. According to the theory of
reduce the resources occupied by the computer, and
information theory, when the distribution of gray values of an
increase the speed of the operation, the color image is first
grayscaled before digital image processing. Generally, the image is relatively average, the amount of information
gray level of the grayscale image is a gray level, and the contained in the image is also large, and the image has a
brightness can be divided into 0 to 255 levels, 0 is the clearer effect from the visual point of the human eye. Median
darkest all black, and 255 is the all white. At present, the filtering technology, median filtering can not only eliminate
most mature technology application is the RGB color mode. the pulse interference noise better, but also effectively reduce
The digital image represented by the RGB mode has three the image edge blur while suppressing the pulse interference.
image components, and the RGB values of three pixels of It is a nonlinear signal processing technique based on the
each pixel respectively reflect the brightness values of the theory of sorting statistics that can effectively suppress noise.
three colors at the pixel. The actual color represented by the It replaces the value of a point in a digital image or a digital
pixel is the result of the color superposition of three sequence with the median value of each point in a
different brightnesses. Since there are 256 kinds of values neighborhood of the point, so that the surrounding pixels are
for each color, there are more than 16 million gray. A pixel with a large difference in degree value is
(256*256*256) color variations per pixel. However, after changed to a value close to the surrounding pixel value, so
conversion to a grayscale image, there are only 256 that an isolated noise point can be eliminated, which is
variations of each pixel, so the amount of computation of effective for salt and pepper noise. The advantage of the
the computer can be greatly reduced. The converted median filter is that it has advantages when filtering out
grayscale image, like the description of the original color superimposed white noise and long tail superimposed noise,
image, still contains the correlation characteristics of the but it is not suitable when there are many details in the image
original image’s chromaticity and brightness [37]. such as points, lines and apex. The improved algorithm has
The purpose of the enhanced technique operation of the the right to median filtering, the switching median filtering
image is to enhance the perceived effect of the image, making algorithm based on the sorting threshold, and the adaptive
it more suitable for a specific application [38]. Purposefully median filter.
highlight certain features of the image, emphasizing the
differences between different images to suit specific D. IMAGE RECOGNITION
situations or special requirements. In a broad sense, as long as In a broad sense, image technology is a general term for
the structural relationship between the parts of the original various image-related technologies. According to the
image is changed, the purpose is to better the application research method and the degree of abstraction, the image

4 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3021590, IEEE Access
Research on Image Recognition Technology Based on Machine Learning

technology can be divided into three levels, which are feature extraction and selection, classifier design and
divided into: image processing, image analysis and image classification decision. The acquisition of information refers
understanding. The technology intersects with computer to the conversion of information such as light or sound into
vision, pattern recognition and computer graphics, and electrical information through sensors. That is to obtain the
biology. Mathematics, physics, electronics, computer basic information of the research object and transform it into
science and other disciplines learn from each other. In information that the machine can recognize by some means.
addition, with the development of computer technology, Preprocessing mainly refers to operations such as de-drying,
further research on image technology is inseparable from smoothing, and transforming in image processing, thereby
theories of neural networks and artificial intelligence. Image enhancing important features of the image. Feature extraction
processing includes image compression, image encoding, and selection means that in pattern recognition, feature
image segmentation, etc. The purpose of processing the
extraction and selection are required. The simple
image is to determine whether the image has the required
understanding is that the images we study are various. If we
information and filter out the noise, and to determine the
need to distinguish them by some method, we must identify
information. Common methods include grayscale,
binarization, sharpening, denoising, etc.; image recognition them by the characteristics of these images. The process of
is to match the processed image, and the category name is acquiring these features is feature extraction. Features
determined. Image recognition can be extracted on the basis obtained in feature extraction may not be useful for this
of segmentation. The features are filtered, and then these recognition. At this time, useful features are extracted, which
features are extracted and finally identified according to the is the choice of features. Feature extraction and selection is
measurement results. The so-called image understanding one of the most critical techniques in the image recognition
refers to the description and interpretation of the image process, so the understanding of this step is the focus of
based on the classification and structure analysis based on image recognition.
image processing and image recognition. Therefore, image On the basis of in-depth learning, image recognition
understanding includes image processing, image technology has been able to recognize moving objects. Its
recognition, and structural analysis. In the image main principle is to process and make decisions on blurred
understanding section, the input is an image and the output image information through intelligent module, and then
is a description of the image. obtain results with high similarity, and then confirm image
The development of image recognition has experienced information through screening. Classical image recognition
three stages: text recognition, digital image processing and model: LeNet is an earlier CNN model (1994). It has three
recognition, and target recognition. Usually, when a domain convolution layers (C1, C3, C5), two pooling layers (S2, S4)
has a requirement that can't be solved by the inherent and one full connection layer (F6). The input image is 32 x
technology, the corresponding new technology will be 32, and the output image is the probability of 0 to 90 digits.
produced. The same is true of image recognition technology. At that time, the error rate of the network model was less than
The invention of this technology is to let the computer instead 1%. LeNet was arguably the first commercially valuable
of human processing a large number of physical information, CNN model since it was successfully used to identify mail
and solve the problem of information that can not be codes. AlexNet is a milestone in the history of CNN
recognized or the recognition rate is very low. Computer development. Compared with LeNet network, AlexNet
image recognition technology is the process of simulating network is not much improved in structure, but has great
human body image recognition. In the process of image advantages in network depth and complexity. AlexNet has
recognition, pattern recognition is essential. Pattern the following meanings. It reveals the powerful learning and
recognition is a basic human intelligence. However, with the expressive ability of CNN, which leads to the upsurge of
development of computer and the rise of artificial CNN research. (2) GPU is used for calculation, which
intelligence, human pattern recognition has been unable to shortens the time and cost of training. (3) Training techniques
meet the needs of life, so human beings hope to replace or such as ReLU activation function, data augmentation and
expand part of human brain labor with computers. This way random inactivation were introduced to provide samples for
the pattern recognition of the computer is created. Simply subsequent CNN.
put, pattern recognition is the classification of data. It is a
science that is closely integrated with mathematics. Most of III. EXPERIMENTS
the ideas used are probability and statistics. Pattern The license plate detection was chosen as the target, the
recognition is mainly divided into three types: statistical study of the target vehicle image recognition method based
pattern recognition, syntax pattern recognition, and fuzzy on machine learning, test software environment: windows
pattern recognition. XP, VC ++ 6.0, Matlab7. The test hardware environment is:
Since computer image recognition technology is the same CPUAMD4400 +, the memory is 2G . The object is to be
as human image recognition, their processes are similar. sorted (3.2.2) H, U, V, W. The training sample comes from
Image recognition technology is also divided into the two aspects: (1) the set of license plate characters after the
following steps: information acquisition, preprocessing, division; (2) the license plate characters obtained by the
manual screenshot. Due to the particularity of the license

VOLUME XX, 2017 5

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3021590, IEEE Access
Research on Image Recognition Technology Based on Machine Learning

plate characters, the amount of characters other than given hypothesis, input: H chromosome. Output: fitness
numbers is small, and it is unrealistic to find a large number value D, fitness calculation as formula.
of test samples that match the actual probability
distribution. Therefore, some characters are added to the 
 pc 1 

 pc 1  pc 2  fb  f 
training sample set by manual means for training. 100 All fb  f
test samples after dividing by the plate image, from the
Pc   
fm  f 

normalized characters. In the process of optimizing the  pc 1 fb  f (3)
weights and thresholds of neural networks, genetic 1
algorithm firstly encodes all weights and thresholds and f 
12 (4)
then synthesizes a long string in a certain order. In this way,
when the network structure has been determined, each After initializing the network with chromosome h, the
chromosome string describes all the parameters of the percentage of incorrect classification in the test set.
neural network, including threshold and weight. In order to According to this fitness function, each hypothesis can be
calculate the fitness of each individual, we assign the evaluated and ranked according to its merits and demerits. In
individual string to the network parameters (decoding). The further evolution, we can select excellent individuals
network calculates the training input samples, and then according to the fitness of each individual and eliminate
returns the sum of squares of errors (fitness) according to inferior individuals.
the output samples. In genetic algorithm, neural network
plays the role of calculating function. Throughout the IV. DISCUSSION
iteration process, the structure of the neural network, The experimental results show that the optimal individual
including the number of nodes in the hidden layer, the has been optimized by the genetic algorithm of 5000
number of nodes and the connection mode between nodes generations. When it evolved to the 3,000th generation, it
are fixed and unchanged. This paper chooses: (1) neural basically converges. The fitness curve is shown in Figure1.
network based on genetic algorithm selection, (2) BP neural
network, (3) KNN algorithm for comparative experiments.
Firstly, according to the formula:
hr io (1)

h  r  i   o  1  1
(2)
Equations 1 and 2 are used to calculate the number of
hidden layers, where R is a random number between 1 and
10, I is the dimension of input layer and O is the dimension
of output layer. Then, the chromosome generated by genetic
algorithm is decoded, and the fitness of the sample in the
test set is tested directly. The adaptive genetic algorithm can
FIGURE 1. Fitness curve
keep the population diversity and ensure the convergence of
The parameters of the BP neural network are: the learning
the genetic algorithm. However, the template based image
matching algorithm requires that the image to be predicted rate is 0.08, and the impulse is 0.1, which is directly trained
and the template must be completely consistent before it can using the training sample set. The accuracy of the training set
be recognized effectively. In practical applications, the finally obtained after training is 0.9756. KNN does not
same image corresponds to complex and changeable feature require training and is an inert classification. Under the
templates in different environments (illumination, assumption that the fitness is Fitness(h1)=0.9543,
deformation, selection, etc.), which requires that machine Fitness(h2)=0.9456, and Fitness(h3)=0.9325, the
learning can not only recognize the image information classification results are shown in Table 1.
characteristics completely consistent with the feature
template, We also need to be able to effectively identify the TABLE I
CLASSIFICATION RESULT
images whose feature templates are not completely Noiseless Noisy Training Recognition
consistent in different environments. When the group Classification
test test time time (ms)
fitness is large, we should choose a smaller cross h1 93.41% 86.90% 12H23M 12
probability. When the fitness is small, a large cross h2 95.82% 85.45% 12H23M 11
probability is used to speed up the search. For example, h3 94.01% 73.56% 12H23M 12
formula 3. Fitness function is an important function (i.e. BP 92.23% 81.96% 12M 12
hypothesis) to evaluate the advantages and disadvantages of KNN 92.67% 88.54% Null value 46
chromosomes. The fitness function constructed in this paper Among the three hypotheses selected, H1, which
is based on the minimum score. Class errors are standard. performs best in training set, does not perform best in
The algorithm flow is as follows: calculate the fitness of the test set. The results show that genetic algorithm can not

6 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3021590, IEEE Access
Research on Image Recognition Technology Based on Machine Learning

completely solve the over fitting problem of neural The overall correct rate in the comprehensive classification
network. In the classification experiment, the overall experiment, when the fitness in this experiment is
accuracy is high. Due to the problems of character offset, Fitness(h2)=0.9456, the performance of the recognition
deformation, skew and blur in the acquired image, any image is better. From the observation of recognition rate, we
text features have spatial distribution. In the case of noise can find that the recognition rate of neural network based on
interference, it is difficult to accurately classify with a genetic algorithm training is better than that based on
few simple features. The license plate recognition results traditional method training. In addition, the traditional KNN
show some differences with different degrees of classification method also shows good results. Although the
adaptation, 2,3,4. recognition time increases with the increase of the number of
training samples, it has good simplicity and anti-interference.
In non-real-time analysis environment, the cross-use of
multiple methods can be considered to further improve the
recognition rate.

V. CONCLUSIONS
As an important method in the field of artificial
intelligence, machine learning has been widely used in traffic
identification research in recent years. Because of its
intelligence, good generalization and high recognition
efficiency, it has gradually become the mainstream of image
recognition research. This paper studies the application of
FIGURE 2. H1 recognition results under fitness image recognition technology based on machine learning in
license plate recognition. In order to complete the research of
this paper, a lot of research on the current development of
license plate recognition research is carried out, and the
horizontal and vertical research and research are carried out
in the field of recognition. Some basic technologies of license
plate recognition are studied, such as image processing,
pattern classification, machine learning, artificial intelligence
and so on.
In order to complete this experiment, a large amount of
target data was collected, but in the field of target
recognition, it is very difficult to obtain large-scale effective
data. This is also the primary problem that hinders the
FIGURE 3. H2 recognition results under fitness application of deep learning in the field of image recognition.
To this end, it is necessary to find a more effective way to
carry out manual data expansion based on the original
database, so that deep learning can be effectively applied.
Data in life is ubiquitous, but tagged data is not common.
Similarly, it is easier to collect data in the field of image
recognition, but manually collecting the collected data is a
time-consuming and labor-intensive task. To this end,
unsupervised learning algorithms are also the focus of
research in deep learning, such as generating confrontational
network models. In the correction process of the license plate,
this paper mainly focuses on the linear information provided
by the framed license plate. If the license plate location
module provides a license plate without a frame, then a
FIGURE 4. H3 recognition results under fitness targeted algorithm should be developed. At the same time, in
view of the control of the generalization accuracy of the
It can be seen from the recognition result that the classifier in the license plate character recognition, this paper
recognition accuracy of Fig. 2 is not high, and the target area combines the genetic algorithm with the optimal solution
cannot be accurately captured. The recognition result in Fig. search tool which is better than the exhaustive method to
3 is good, and all the target areas in the image can be solve the global space of the weight of the neural network.
captured, and the target area in Fig. 4 identifies the defect. After experimental verification, the three solutions with the

VOLUME XX, 2017 7

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3021590, IEEE Access
Research on Image Recognition Technology Based on Machine Learning

highest fitness are obtained from the genetic algorithm. The preprocessing for DRAM-PCM Hybrid Main Memory”, IEEE
Computer Architecture Letters, 2018, 17(2):163-166.
generalization effect after decoding to the neural network is [20] Zare F, Ansari S, Najarian K, et al. Preprocessing Sequence
relatively good. Coverage Data for Precise Detection of Copy Number Variations”,
IEEE/ACM Transactions on Computational Biology and
ACKNOWLEDGEMENTS Bioinformatics, 2018, PP(99):1-1.
[21] Lei T, Jia X, Zhang Y, et al. Holoscopic 3D Micro-Gesture
This work was supported by Key scientific research project Recognition Based on Fast Preprocessing and Deep Learning
of Neijiang Normal University (NO.18ZA06); Natural Techniques[C]// 2018:795-801.
Science Foundation of Zhejiang Province. [22] Moravčík M, Schmid M, Burch N, et al. DeepStack: Expert-level
artificial intelligence in heads-up no-limit poker”, Science, 2017,
356(6337):508.
REFERENCES [23] Wang X, Li X, Leung V C M. Artificial Intelligence-Based
[1] Shi X, Chen Z, Wang H, et al. Convolutional LSTM Network: A Techniques for Emerging Heterogeneous Network: State of the Arts,
Machine Learning Approach for Precipitation Nowcasting”, Opportunities, and Challenges”, IEEE Access, 2017, 3:1379-1391.
2015:961-997. [24] Gaber, T., Abdelwahab, S., Elhoseny, M., Hassanien, A.E., Trust-
[2] Xiao H, Rasul K, Vollgraf R. Fashion-MNIST: a Novel Image based secure clustering in WSN-based intelligent transportation
Dataset for Benchmarking Machine Learning Algorithms”, systems, Computer Networks 146, pp. 151-158, 2018
2017:1691-1737. [25] Mukhtar A, Xia L, Tong B T. Vehicle Detection Techniques for
[3] Raccuglia P, Elbert K C, Adler P D F, et al. Machine-learning- Collision Avoidance Systems: A Review”, IEEE Transactions on
assisted materials discovery using failed experiments”, Nature, Intelligent Transportation Systems, 2015, 16(5):2318-2338.
2016, 533(7601):73. [26] Razakarivony S, Jurie F. Vehicle detection in aerial imagery : A
[4] Baydin A G, Pearlmutter B A, Radul A A, et al. Automatic
small target detection benchmark ☆”, Journal of Visual
differentiation in machine learning: a survey”, Computer Science,
Communication & Image Representation, 2015, 34:187-203.
2015(February):451-479.
[27] Chen X, Xiang S, Liu C L, et al. Vehicle Detection in Satellite
[5] Giusti A, Guzzi J, Dan C C, et al. A Machine Learning Approach to
Images by Hybrid Deep Convolutional Neural Networks”, IEEE
Visual Perception of Forest Trails for Mobile Robots”, IEEE
Geoscience & Remote Sensing Letters, 2017, 11(10):1797-1801.
Robotics & Automation Letters, 2017, 1(2):661-667.
[28] Rezaei M, Terauchi M, Klette R. Robust Vehicle Detection and
[6] Couprie M, Bezerra F N, Bertrand G. Topological operators for
Distance Estimation Under Challenging Lighting Conditions”, IEEE
grayscale image processing”, Journal of Electronic Imaging, 2015,
Transactions on Intelligent Transportation Systems, 2015,
10(10):1003-1015.
[7] Mullapudi R T, Vasista V, Bondhugula U. PolyMage:Automatic 16(5):2723-2743.
Optimization for Image Processing Pipelines”, Acm Sigarch [29] Reddy, M. R., Srinivasa, K. G., and Reddy, B. E. 2018. "Smart
Computer Architecture News, 2015, 43(1):429-443. Vehicular System Based on the Internet of Things," Journal of
[8] Guido Dartmann, Houbing Song, and Anke Schmeink. Big Data Organizational and End User Computing (30:3), pp. 45-62.
Analytics for Cyber-Physical Systems: Machine Learning for the [30] Elhoseny, M., Yuan, X., El-Minir, H.K., Riad, A.M., Extending self-
Internet of Things. ISBN: 9780128166376. Elsevier, 2019, pp. 1- organizing network availability using genetic algorithm, 5th
360. International Conference on Computing Communication and
[9] Zhang Y, Kwong S, Wang X, et al. Machine learning-based coding Networking Technologies, ICCCNT 2014, (DOI:
unit depth decisions for flexible complexity allocation in high 10.1109/ICCCNT.2014.6963059)
efficiency video coding.”, IEEE Transactions on Image Processing, [31] Wang, Q., & Lu, P. (2019) “Research on Application of Artificial
2015, 24(7):2225-2238. Intelligence in Computer Network Technology”, International
[10] Wang B., Chen L.L.(2019). Novel image segmentation method Journal of Pattern Recognition and Artificial Intelligence, 33(5),
based on PCNN, Optik, 187, 193-197. 1959015.
[11] Wang K, Zhang D, Li Y, et al. Cost-Effective Active Learning for [32] Bangzhu, Z., Shunxin, Y., Minxing, J. (2019) “Achieving the
Deep Image Classification”, IEEE Transactions on Circuits & Carbon Intensity Target of China: A Least Squares Support Vector
Systems for Video Technology, 2017, PP(99):1-1. Machine with Mixture Kernel Function Approach”, APPLIED
[12] Yuan Y, Lin J, Wang Q. Hyperspectral Image Classification via ENERGY, 233, pp. 196-207.
Multitask Joint Sparse Representation and Stepwise MRF [33] Ibrahim M. El-Hasnony ; Sherif Barakat ; Mohamed Elhoseny ;
Optimization.”, IEEE Transactions on Cybernetics, 2017, Reham R. Mostafa, Improved Feature Selection Model for Big Data
46(12):2966-2977. Analytics, IEEE Access, Vol 8, No 1, PP: 66989-67004, 2020 (DOI:
[13] Cheng F, Hong Z, Fan W, et al. Image Recognition Technology 10.1109/ACCESS.2020.2986232)
Based on Deep Learning”, Wireless Personal Communications, [34] Sachi Nandan Mohanty, E. Laxmi Lydia, Mohamed Elhoseny,
2018(C):1-17. Majid M. Gethami Al Otaibi, K. Shankar, Deep learning with LSTM
[14] Lin B S, Liu C F, Cheng C J, et al. Development of Novel Hearing based distributed data mining model for energy efficient wireless
Aids by Using Image Recognition Technology”, IEEE Journal of sensor networks, Physical Communication, 2020, In Press
Biomedical & Health Informatics, 2018, PP(99):1-1. (DOI:https://doi.org/10.1016/j.phycom.2020.101097)
[15] Zhang X B, Ge X G, Jin Y, et al. [Application of image recognition [35] Elsayed, W., Elhoseny, M., Sabbeh, S., & Riad, A. (2018). Self-
technology in census of national traditional Chinese medicine maintenance model for wireless sensor networks. Computers &
resources]”, Zhongguo Zhong yao za zhi = Zhongguo zhongyao Electrical Engineering, 70, 799-812.
zazhi = China journal of Chinese materia medica, 2017, [36] S Wan, Y Xia, L Qi, YH Yang, M Atiquzzaman. Automated
42(22):4266. colorization of a grayscale image with seed points propagation.
[16] Sun D, Gao A, Liu M, et al. Study of real-time detection of bedload IEEE Transactions on Multimedia, 2020.
transport rate using image recognition technology”, Journal of [37] Xiong, Z., Wu, Y., Ye, C., Zhang, X., & Xu, F. (2019). Color image
Hydroelectric Engineering, 2015, 34(9):85-91. chaos encryption algorithm combining CRC and nine palace map.
[17] Fukada H, Kasai K, Shou O. A Field Experiment of System to Multimedia Tools and Applications, 78(22), pp.31035-31055.
Provide Tourism Information Using Image Recognition Type AR [38] He, Y., Xiong, W., Chen, H., Cao, C., Huang, W., Yang, L.,... Chen,
Technology”, Lecture Notes in Electrical Engineering, 2015, Y. (2019). Image quality enhanced recognition of laser cavity based
312:381-387. on improved random hough transform. Journal of Visual
[18] Wang, B., Chen, L., Zhang, Z.(2019). A novel method on the edge Communication and Image Representation, 102679. doi:
detection of infrared image, OPTIK, 180, 610-614. 10.1016/j.jvcir.2019.10267
[19] Yun J T, Yoon S K, Kim J G, et al. Regression prefetcher with

8 VOLUME XX, 2017

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.3021590, IEEE Access
Research on Image Recognition Technology Based on Machine Learning

YANPING WANG was born in Hangzhou, Zhejiang, P.R. China, in 1969.

He received the bachelor‘s degree from Hangzhou Dianzhi University,
P.R. China, and received the MBA from Tongji University, P.R. China
.Now, he is work in Information Center of China Tobacco Zhejiang
Industry Co. Ltd. His research interest include computational intelligence,
network security, machine vision and video analysis.
E-mail: [email protected]

LIJUAN LIU was born in Linfen, Shanxi. P.R. China, in 1975. She
received the master‘s degree from University of Electronic Science and
Technology of China, P.R. China. She works in College of Computer
Science, Neijiang Normal University. Now, She is studying for a doctorate
in intelligent information integration at Makuwon University in South
Korea. Her research interest include data mining, intelligent fusion and big
data analysis.
E-mail: [email protected]
WANLE CHI was born in Ruian, Zhejiang, P.R.
China, in 1978. He received the bachelor‘s degree from Dalian University
of Technology and master's degree from Zhejiang University. Now, he
studies in Faculty of Information and Communication Technology,
Technical University of Malaysia Malacca, Malaysia for PHD. degree , and
works in Department of Information Technology, Wenzhou Polytechnic.
His research interest include pattern recognition, machine learning, and big
data analysis.
E-mail: [email protected]

VOLUME XX, 2017 9

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.