https://doi.org/10.31449/inf.v48i5.5345                                              Informatica 48 (2024) 29–40  29 
Research on Automatic Recognition Technology of Library Books 
Based on Image Processing 
Haiyan Xun
 
Library of Shandong Women's University, Jinan, Shandong, 250300, China 
E-mail: xunhaiyan2023@126.com 
Keywords: image processing, automatic recognition technology, image retrieval 
Received: 
The intelligence of computers is the future development direction. In today's society, the amount of 
information is increasing, which puts forward higher requirements for retrieval technology and 
automation levels. With the development and popularization of the Internet era, online shopping, study, 
and even work have become the norm in people's lives. However, a large amount of data is generated on 
the Internet every day, and how to obtain the information people need from this data is a key research 
problem today. It can be seen that the traditional global search or image search can no longer meet the 
amount of information people need, so the content-based search method will inevitably become a more 
popular database retrieval technology. In recent years, content-based image capture has become a 
research center in the field of image information removal. This article first studies image processing and 
capture technology. It provides a detailed overview of image segmentation technology and image 
segmentation models introduced in image processing, as well as three image segmentation methods. This 
article also introduces the basic principles and framework of the CBIR procurement system. The most 
important technology in the CBIR system is the analysis and description of image features, and several 
methods commonly used to express the content of feature images. Secondly, it analyzes the image edge 
detection algorithm in detail and finally introduces the functional division and workflow of the library 
book automatic recognition system. It also provides all the construction environments, processes, and 
algorithms to perform the basic functions of the system. 
Povzetek: Predstavljena je tehnologije obdelave in zajemanja slik, podrobno še segmentacija slik in 
metode segmentacije. 
 
1   Introduction 
In modern scientific research and daily life, people 
usually prefer to obtain information through images. 
However, traditional manual identification has been 
unable to meet the needs of social development for the 
research on the automatic identification technology of 
artificial intelligence library books. With the 
development of various computer technologies and 
networks, the informatization development of various 
enterprises has become the current mainstream 
development direction. In this information age, people 
usually learn about social changes through website 
information. The library is the dissemination center of all 
kinds of information in colleges and universities [1-2]. 
With the popularization of information technology, major 
libraries have also established online libraries. Such 
online libraries can not only effectively manage the 
books in the library, but also facilitate people to read 
books. At the same time, the online library also speeds up 
the time for people to find the books they need. However, 
due to the large number of secretaries in the library, the 
order of books in the online library system is also entered 
according to the order of the books in the offline library. 
Therefore, there may be errors in the information of the 
book in the system due to the incorrect placement of the 
book. This has a very bad influence on the development 
of online libraries [3]. 
Image segmentation usually refers to the segmentation of 
digital images. The method of image segmentation is the 
process of dividing an image into several discontinuous 
regions. The basic idea is to classify and group images 
with similar feature values in digital images, and divide 
the images into regions with different important levels, 
thereby reducing the amount of data and information 
contained in each part of the image. The structure 
information is suitable for the later stage of image 
processing, and it is further processed to store 
information about the target structure while significantly 
reducing the amount of target data [4-5]. 
Image segmentation techniques can be divided into three 
categories. One is based on the extracted regional 
features or the morphological results obtained from the 
regional features. It uses thresholds, regional methods, 
and textures to classify image pixel groups; the second is 
to pay attention to the picture's border, which divides the 
margin to obtain the region according to the obtained 
operator. The last method is to use statistical features and 

30   Informatica 48 (2024) 29–40                                                                       H. Xun 
prior knowledge features for image segmentation. This 
method is to segment first, then highlight the objects, and 
then combine them into segmented regions according to 
the domain of knowledge [6]. 
According to the three categories of image segmentation, 
the most commonly used segmentation techniques are the 
threshold method, edge detection method, and region 
improvement method. In threshold segmentation, the 
maximum threshold and blur threshold are more suitable 
for the segmentation of specific types of images (vessels) 
[7-8]. Although that which separates the background 
from the target is not very obvious, the segmentation 
effect is not very good; the method of edge detection is to 
first detect the pixels along the edge of the image, and 
then stitch the end pixels together to form a segmented 
area. The edge detection method uses part of the window 
operation to find the edge information in the image. This 
method is suitable for images with a solid color 
background [9-10].Literature survey are discussed in 
Table 1. 
 
Table 1: Related works 
Ref. 
No 
Methodologies Summary of findings Limitation 
[11] To facilitate effective 
deep-learning medical image 
processing, the study presents 
TorchIO, a Python package. 
Patch-based sampling, spatial 
metadata consideration, and 
data augmentation are 
highlighted as ways to 
overcome the difficulties 
associated with processing 
MRI and CT images 
Patch-based sampling, augmentation, 
preprocessing, and loading are made 
easier by TorchIO. It supports 
composition, transform inversion, and 
simulation of MRI-specific artifacts. It 
interfaces with PyTorch and 
conventional medical image processing 
libraries. Open-science principles are 
encouraged by the availability of the 
source code, tutorials, and 
documentation. 
The efficiency of TorchIO, 
which simplifies medical 
image processing, could 
differ depending on the 
application case. The 
modularity and 
interoperability of the 
library with various 
deep-learning frameworks 
for medical pictures should 
be taken into consideration 
by users. 
[12] To extract gear tooth profiles 
properly without depending on 
conventional meshing theory, 
the research presents a novel 
method called 
Engagement-Pixel Image Edge 
Tracking (EPIET). The 
process entails extracting 
meshing points, calibrating 
tool locus coordinates, and 
obtaining immediate contact 
images. 
The method's efficiency in removing 
tooth profile edges is illustrated by a 
case study on a face gear. Its promise 
for computerized design of complicated 
conjugate curved surfaces is 
demonstrated by the results, which 
show feasible accuracy and stability 
when compared to standard meshing 
equations. 
The paper discusses major 
error sources associated 
with the presented method. 
While it shows promise for 
gear profile extraction, 
further research is needed 
to explore its applicability 
to diverse gear types and 
potential limitations in 
complex scenarios. 
[13] Using a case study 
methodology, this paper 
examines how public libraries, 
namely the Seattle Public 
Library (SPL) system, curate 
and use available demographic 
data. As part of the study, use 
cases are created, a dashboard 
tool prototype is developed 
using available census data, 
and information needs are 
identified through interviews 
with SPL regional managers. 
 
The study provides new perspectives on 
how open data might be used to help 
SPL regional managers find the 
information they need. The needs 
within two SPL regions are successfully 
addressed via a dashboard tool 
prototype. As a result of the findings, 
public libraries may find it easier to 
keep up with changing neighborhood 
demographics by developing 
reproducible data analytic techniques. 
Limitations include the 
case study's specialization 
to the SPL system, which 
may restrict 
generalizability even 
though the paper provides 
insightful information. 
Furthermore, issues 
concerning the timeliness 
and accuracy of publicly 
available demographic data 
may affect how broadly 
applicable these approaches 
are in various public library 
environments. 
[14] The goal of the project is to 
increase the efficiency of 
library book borrowing by 
developing an automated book 
sorter that makes use of RFID 
The device’s great precision and quick 
reaction time when processing book 
numbers and operating equipment are 
demonstrated through practical 
operation. Effective book classification 
Even though the automated 
book sorter works well, 
there could be issues with 
managing different book 
sizes or fixing system 

Research on Automatic Recognition Technology of Library Books…                      Informatica 48 (2024) 29–40   31 
technology and a single-chip 
microprocessor control 
system. The sorter uses an AC 
motor to power a conveyor 
belt, recognizes electronic tags 
using RFID and a 
microcontroller, and precisely 
places books in recycling bins. 
is made possible by the combination of 
RFID technology and a single-chip 
microcomputer, which improves library 
operations and improves student 
learning outcomes. 
errors. It might take 
constant improvement and 
adjustment to maximize 
performance in different 
library circumstances. 
[15] Using Particle Image 
Velocimetry (PIV) to assess 
multiphase flows, the study 
highlights the importance of 
appropriate image 
segmentation for precise phase 
dynamics separation. 
Triangular meshing and 
particle detection are used in 
the suggested approach to 
provide reliable phase 
separation and interface 
detection. 
The new technique uses a 2D 
unstructured mesh for phase separation 
and interface detection and successfully 
identifies tracer particles by utilizing 
seeding density differentiation between 
phases. The effectiveness of the method 
is demonstrated by parametric analysis 
on synthetic images, and successful 
tracking of complicated interface 
evolution is revealed by experimental 
application in immiscible multiphase 
flow in porous media. 
While there is promise in 
the method, more 
validation under various 
experimental conditions is 
required to verify its wider 
application, as potential 
constraints may develop in 
other flow scenarios. 
 
 
1.1  Problem statement 
There are problems with misplacing books since the 
current library book management systems are not effective 
in identifying books. By using image processing 
technologies, this research proposes an automatic 
identification technique to address the issue. Accurate 
character recognition, post-processing analysis, and 
bookcase detection are the system's main goals. 
Improvements in orderliness, real-time book tracking, and 
user-friendly reading environments are all part of the plan 
to improve library management. 
2  Research on image segmentation 
technology and retrieval 
technology 
2.1 Image segmentation technology 
General model of image segmentation 
According to the basic concept of image segmentation, 
the known image segmentation is the process of dividing 
an image into several regions. This segmented area is a 
combination of pixels with common numerical features. 
For example, different objects occupy different areas of 
the image, and the image and background objects occupy 
different areas of the image. The segmented regions must 
have sufficient homogeneity and connectivity to be 
segmented together[8].  
 
 
Let F be the collection of all the image's pixels G, and the 
hypothesis about the uniformity of 

n
j
j
F S
1 =
= is P (.)[3]. 
The mathematical explanation of the above situation is 
shown by formulas (1), (2), (3), (4): 

n
j
j
F S
1 =
=                (1) 
) ( , j i S S
j i
  =        (2) 
) ( , TRUE ) ( j S P
j
 =
      
 (3) 
) ( , ) ( j i FALUS S S P
j i
 =  (4) 
Summary of image segmentation methods 
(1) Threshold segmentation 
Threshold segmentation can be divided into three 
methods: global threshold method, multi-threshold 
method, and adaptive threshold method. Among them, the 
global threshold method is easier to apply to complete 
images with obvious contrast between the target and the 
background, especially if the level of the gray 
background is fixed, the effect is much better.  
 
 
 
 

32   Informatica 48 (2024) 29–40                                                                       H. Xun 
The multi-threshold method is suitable for images with 
different target types and background areas, and different 
thresholds can achieve different targets. The main core 
meaning of multi-threshold is to set multiple thresholds 
so that for images with different target types and 
background areas, multiple grayscales in each target and 
background area in the image can be compared, which 
enables more accurate detection. The gray value of the 
pixel, and more accurately segment the image. The 
adaptive threshold will be adjusted according to the 
change in the gray background level, and the background 
from the target will also change. If the threshold is fixed, 
the target effect of this kind of image acquisition is not 
good, and the adaptive threshold method, that is, the gray 
threshold obtained under different image conditions is 
also different, and it may have a better segmentation 
effect [9]. 
The essence of image segmentation is to split a complete 
image into multiple small images, and the basis for 
splitting into small images is based on the feature 
attributes of each pixel in the image, the pixel points that 
are closer and have similar feature attributes are grouped 
into a small image, and then these pixel points are used as 
the center to expand outward to find similar pixel points, 
and if similar ones are found, they can be fused into the 
small image, and so on until the whole image is split. 
The global threshold method is suitable for the complete 
image with obvious contrast between the target and the 
background, but it has certain disadvantages when it is 
used to segment the image where the contrast between the 
target and the background is not obvious and incomplete. 
The disadvantage of incomplete images is that some of 
the pixel point properties in the image are changed or lost, 
the histogram of these incomplete images is generally 
close to a single peak, and the values derived from 
traditional thresholding are generally not at the bottom of 
the valley in the histogram, so if the incomplete images 
are then segmented using traditional thresholding 
segmentation methods cannot get the correct threshold 
values and thus also cannot segment the images 
reasonably.  
When the image threshold is segmented, the 
segmentation of the target and the background is based on 
comparing the gray pixel value of the position with the 
specified threshold. The classic method of selecting a 
threshold is the maximum between-class variance method, 
which is used alone to segment a single threshold. The 
principle is to maximize the difference between the two 
parts of the classroom by setting a threshold.  
 
(2) Edge detection 
The edge is one of the important components of the 
image, so edge detection must be performed when image 
segmentation. The edge of the image is generally between 
the target image and the background. If the edge 
information in the image can be obtained during image 
segmentation, this can improve the efficiency of image 
segmentation [10]. 
The gradient operators of the commonly used edge 
detection algorithms include the Robert operator, Sobel 
operator, Prewitt operator, Laplace operator, etc. 
Roberts’s edge detection operator has two pairs of 
detection operators, which are the detection operators in 
the vertical direction and the diagonal direction. These 
two pairs of detection operators detect the difference 
between two adjacent pixels in their respective directions 
at the same time. It can be seen that the Roberts edge 
detection operator can detect the edges of the image more 
accurately, but the Roberts edge detection operator is 
more sensitive to the noise generated in image 
segmentation, so it is not suitable for image segmentation 
with blurred edges and large segmentation noise [11]. 
The principle of the Roberts edge detection operator is to 
use the vertical and diagonal detection operators to detect 
the difference between two adjacent pixels in two 
directions in an image. The characteristics of the Roberts 
edge detection operator are as follows: Roberts is more 
sensitive to the noise generated by image segmentation, 
which makes this operator not suitable for image 
segmentation with fuzzy edges and more complicated 
segmentation. However, noise will inevitably occur in the 
image segmentation process, so there is a certain error in 
the edge of the image detected by this operator. 
The biggest advantage of the Sobel operator is that it has 
a certain ability to suppress noise when performing edge 
detection. This ability enables the Sobel operator to help 
the image segmentation more smoothly, which makes the 
edges of the image segmentation clearer, and it can also 
effectively remove false edges. The traditional Sobel edge 
detection operator's edge positioning and noise smoothing 
are contradictory. To overcome this shortcoming, people 
use edge detection combined with template matching to 
effectively adjust this contradiction. 
Sobel operator principle: In image processing, the 
difference between adjacent pixels or spaces is often used 
to represent image edge information. The principle of the 
Sobel operator is to use the corresponding image data 
convolution mode to perform weighted calculation and 
approximate calculation on discrete data. It uses the small 
convolution mode to perform horizontal edge detection 
and vertical border recognition. 
The edge location of the Sobel edge detection operator is 
contradictory to noise smoothing. To overcome this 
shortcoming, people use edge detection combined with 
template matching to effectively adjust this contradiction. 
The first is to increase the inclusion of the Sobel operator 
relative to the edge direction. Due to the different edge 
directions, 6 templates with different edges moving 45 
degrees clockwise can be added [12]. Secondly, the Sobel 
operator algorithm is improved, that is, the result of the S 
convolution operation on the 8 modes M1-M8: S1 = 
a1+2a8+a7-a3-2a4-a5. So Si (1 = <i = <8)), and finally S 
= max {Si} (1 = <i = <8). 
The traditional calculation method of the Prewitt operator 
is: first do convolution, then compare the dimensions, and 
use the result of the convolution operation as the final 

Research on Automatic Recognition Technology of Library Books…                      Informatica 48 (2024) 29–40   33 
value. 7 convolution operations are unnecessary, resulting 
in a large amount of calculation. This amount of 
calculation is too disadvantageous for processing images 
quickly in real-time. 
There are 7 extra convolution operations under the 
conventional calculation method of the Prewitt operator, 
which causes a large amount of calculation. This is very 
disadvantageous for fast real-time processing of images. 
In response to this problem, the researchers considered 
minimizing the convolution operation, thereby reducing 
the number of operations. If the comparison operation is 
performed first to maximize the convolution mode, and 
then the convolution is performed as expected, then if the 
convolution operation is performed only once, the 
amount of calculation will be greatly reduced, thereby 
improving the computational efficiency. 
The Laplace operator is a second-order differential, which 
doubles the influence of noise, resulting in sharp edges in 
the image. Its advantages are isotropy, that is, rotation 
invariance and displacement invariance. It can be known 
from differential calculus that only linear combination 
operators composed of even-order derivatives and 
even-order odd-order derivatives must be isotropic. 
Image Retrieval Related Technologies ： CBIR 
technology can be traced back to 1992. This technology 
is mainly used to solve the problem of excessive 
newspaper image capture. CBIR is a selection method 
that directly uses image content as image information to 
request an image. The main techniques involved are: 
highlighting features, comparing similarities, and 
extracting matching results. For CBIR, there are many 
algorithms, and the basic extraction methods include 
extraction methods based on color features, texture 
features, shape features, and spatial features [13-14]. 
Based on the example of image query, the CBIR system 
framework can be divided into three modules: function 
search, function comparison, and result display. The 
system framework is shown in Figure 1 below. 
 
Figure 1: The system framework of the CBIR system 
3  Image edge detection algorithm 
Mathematical Background ： The definition of the first 
derivative of the image grayscale in the digital picture is 
shown in formula (5), and the expression of the second 
derivative is shown in formula (6): 
) ( ) 1 ( ) ( ) ( ) ( x f x f x f x x f x f − + = −  + =  (5) 
) ( ) 1 ( 2 ) 2 ( ) ( ) 1 ( ) ( x f x f x f x f x f x f + + − + =  − +  =   (6) 
Edge Model ： When selecting the original features of an 
image, edges are important features in image analysis. 
They are located at the boundaries of different regions 
and represent significant local changes in image intensity. 
These changes appear as discontinuities in image 
intensity, disc discontinuities, or line discontinuities. As 
shown in Figure 2, phase discontinuity refers to the 
sudden change of image intensity from one initial value 
to another, and line discontinuity refers to the sudden 
change of image intensity from one initial value to 
another, and then restores to the initial value. In reality, 
the low-pass filtering function is introduced by the sensor, 
and the two discontinuities are shown as slope edges and 
roof edges [15]. 
 
Figure 2: Edge model diagram 
3.1 Canny edge detection operator 
There are two main types of basic linear recognition 
methods for raster images. One is based on a grayscale 
equation containing additional conditions (such as 
gradient), and the other is based on the edge of the object. 
In most applications, the purpose of detecting a straight 
line is to determine the object in the image, and the object 
should be bounded at the boundary. A boundary is a 
special pixel that must meet a certain boundary when 
creating a target edge. Although it is directly based on 
grayscale and its transformation, it is usually not tested 
whether it meets the edge properties, and the observed 
lines may not be the actual edges that are needed. This is 
the direct reason why related algorithms are generally 
considered wrong. For this reason, it is more convenient 
to perform line detection based on edge analysis [16-17]. 
Among many edge detection algorithms, the canny 
algorithm generally uses a fixed threshold for edge 
detection. Of course, the Canny algorithm generally has 
two thresholds. When using this algorithm to segment the 
target image and the image with more background, its 
efficiency is not high. In addition, some local edges with 
slow changes in gray value will be lost, so this algorithm 
may also have the defect of false edges or edge loss. The 
canny operator is recognized as an operator with a low 

34   Informatica 48 (2024) 29–40                                                                       H. Xun 
error rate, accurate positioning, and strong 
noise-elimination ability. It can make more accurate 
decisions on edge pixels, and it is widely used in edge 
detection. The edge analysis algorithm used by this 
algorithm is executed as follows: let f (x, y) be the input 
gray image, and G (x, y) be the two-dimensional 
Gaussian function. Canny's edge detection algorithm first 
performs audio filtering and Gaussian filtering on the 
input image to find the best value between the edge 
positions. According to formula (7), a well-balanced and 
uniform image is obtained, in which the two-dimensional 
Gaussian function is expressed by formula (8): 
) , ( * ) , ( ) , ( y x f y x G y x F = (7) 
2
2 2
2
2
2
1
) , (


y x
e y x G
+
−
=
   
 (8) 
Gaussian filtering is performed on the image, that is, the 
specified o, x, and y are taken to obtain the corresponding 
Gaussian kernel, which is convolved with the image [12]. 
On this basis, the partial derivative is used to calculate the 
gradient amplitude and gradient direction as shown in 
formulas (9) and (10): 
2 2
| ) , ( | ) , (
y
F
x
F
y x F y x M


+


=  =  (9) 
| ) , ( | / ) , ( y x F y x F n   =           (10) 
Among them, the calculation of the partial derivative can 
be obtained by the first-order difference approximation. 
The gradient operator used by the canny operator is 
shown in formulas (11), (12), and (13): 






=






=
1    1  
1 -  1 -
,
1  1 -
1  1 -
y x
C C      (11) 
)) , 1 ( ) 1 , 1 ( ) , ( ) 1 , ( (
2
1
y x F y x F y x F y x F
x
F
+ − + + + − + =


      
(12) 
)) 1 , ( ) 1 , 1 ( ) , ( ) , 1 ( (
2
1
+ − + + + − + =


y x F y x F y x F y x F
y
F
   (13) 
Because canny edge detection requires a single edge 
response, the principle of non-maximum suppression (as 
shown in Figure 3) is introduced to refine the edges of the 
image. 
 
Figure 3: Non-maximum suppression 
To obtain more accurate edge results, the dual-threshold 
method is used to filter noise below the lower threshold 
and maintain strong edges above the higher threshold. 
Then, after the cropping operation, the gaps between the 
strong corners are connected to obtain a convenient edge 
with a pixel angle. The double threshold method is as 
follows: 
Set the upper limit to TH and the lower limit to TL. 
Usually, TH/TL is between 2-3. Using the two boundaries 
of the gradient magnitude image M(x, y), the strong edge 
is obtained according to formula (14), and the weak edge 
is obtained according to formula (15): 
H H
T y x M y x M  = ) , ( ) , (
          
(14) 
L H L
T y x M y x M y x M  − = ) , ( ) , ( ) , ( (15) 
The Canny algorithm generally has two thresholds, and 
its efficiency is not high when it is used to segment the 
target image and the image with more background. The 
core reason for this problem is that the traditional Canny 
algorithm uses a fixed threshold when detecting the edge 
of the image. This leads to the lack of flexibility and 
adaptability of the traditional Canny algorithm. Therefore, 
researchers have added a genetic algorithm to the 
traditional Canny algorithm. The Canny algorithm based 
on the genetic algorithm can adaptively generate a 
dynamic threshold according to the complexity of the 
image and the change of the gray value, thereby 
effectively solving the problem of low image 
segmentation efficiency. 
The traditional Canny edge detection algorithm causes 
the problem of image segmentation due to the problem of 
fixed threshold, which may cause false edges or missing 
edges with slow changes in gray value. This is very 
unfavorable for image edge detection and acquisition. In 
response to this problem, researchers have proposed a 
method that combines statistical filters and gray-scale 
iteration to calculate the threshold.  
 
 
 
 
 

Research on Automatic Recognition Technology of Library Books…                      Informatica 48 (2024) 29–40   35 
This can not only suppress noise through statistical filters 
but also ensure that edge detection is not affected by 
noise; secondly, the threshold is determined by gray-scale 
iteration calculation, which avoids the problem of edges 
that are lost due to slow gray-scale changes. 
The canny operator is an edge detection algorithm. Its 
specific steps are as follows: First, the image is smoothed 
by Gaussian filtering to reduce the noise of image 
segmentation. Secondly, use relevant calculation formulas 
to find the size and azimuth of the image gray gradient. 
Then the non-maximum value suppression is performed 
on the gradient amplitude obtained in the gradient 
direction, and the local maximum point is found. Finally, 
upper and lower thresholds are used to detect and connect 
edges. 
 
3.2 Image edge tracking 
After the image is divided into several areas, the 
computer will usually visualize and describe the set of 
pixels it contains suitably for further processing. To 
obtain the information of a region, it is usually 
represented by an outer boundary or a group of inner 
pixels. According to the linear characteristics of the grid, 
this paper uses Freeman's chain code as the edge 
description and introduces a fault-tolerant mechanism to 
track the straight line pattern, and then detect the straight 
line [18-19]. 
 
1. Freeman chain code 
The serial code is used to indicate a boundary composed 
of straight-line segments connected in a specified 
direction and length. Usually based on 4 adjacent pixels 
or 8 adjacent pixels, each connection address uses a 
digital code, as shown in Figure 4, it is a Freeman series 
code. This article uses this structure to describe the 
subsequent results. 
 
Figure 4: Freeman chain code 
2. The characteristics of the grating line 
To draw quickly, people have carried out extensive 
research on straight-line grids. Any straight-line segment 
can only be composed of adjacent straight-line segments 
with a specific pattern. The "pattern" described here 
refers to a short straight line that has the same length or is 
separated by only one pixel. Raster line pixels have 
strong directivity, see Figure 5. 
 
Figure 5: The characteristics and direction of the grating 
line 
3. Fault tolerance mechanism 
Under unfavorable conditions, if the linear direction and 
pattern of the grating are strictly used for pixel occlusion 
and tracking, many line segments will be damaged. These 
small line segments are not easy to connect due to the 
large gap. Therefore, a specific fault tolerance mechanism 
must be introduced in the subsequent process. As shown 
in Figure 6, in the constraint mode, the raster line 
tracking algorithm cannot track the entire target at one 
time. Therefore, the whole line can be realized through 
post-processing such as segmentation, capture, and 
polyline. However, this document introduces a 
fault-tolerant mechanism in the tracking process, which 
allows tracking in three directions on one side and 
accommodates a single outlier to achieve a more 
complete straight-line pattern. 
 
Figure 6: Straight-line tracking under non-ideal 
conditions 
In addition, to achieve a complete line segment in one 
round of tracking and improve the efficiency of the 
algorithm, as shown in Figure 7, the line detection 
algorithm proposed in this paper starts to track the edges 
on both sides in six directions. Only two processes of 
edge extraction and straight-line tracking are required. 
These two processes do not require post-processing 
operations such as segmentation, capture, and connection 
in traditional edge point tracking algorithms. 

36   Informatica 48 (2024) 29–40                                                                       H. Xun 
 
Figure 7: Tracking in six directions on both sides 
4. Result and discussion 
4.1 Library book automatic identification 
system 
System design scheme 
1. Function description 
As shown in Figure 8, the library book automatic 
recognition system is mainly composed of three modules: 
1) Preprocessing module, which processes the image of 
the spine on the bookshelf from the outside. The main 
functions include trunk segmentation, label, telephone 
number, and character number extraction and 
segmentation;2) The recognition module is mainly used 
to recognize the character images in the preprocessing 
link, and then recognize the books on the shelf;3) 
Post-processing functions include result identification 
(clutter analysis), database interaction analysis, and result 
feedback. The pre-processing module is a server that 
processes the spine images collected by the administrator. 
The processing of the LFD proposed in the article is 
based on a line detection algorithm, which is used to 
detect the edge of the spine of a book; Extracting the 
phone number label from the label in the HSV space ROI 
is based on the characteristics of the label (such as aspect 
ratio, area, etc.); Phone number character segmentation, 
according to the projection method, it divides the phone 
number string in the ROI into a single character image 
sequence, which is used in the system recognition 
process. 
In the recognition module, the server recognizes the 
character sequence of the phone number of each book by 
calling a pre-trained deep learning model (feedforward 
double convolutional network) and saves the recognition 
results for further processing. 
The post-processing module analyzes and summarizes the 
past verification results. The library is an analysis of the 
verification results by interacting with the database 
management system. The system proposed in this paper 
currently only focuses on the recognition of books in the 
wrong order, which can improve the intelligent borrowing 
and reading functions. In addition, it is also important to 
provide feedback on the analysis results, and guide and 
assist the management staff in their work. 
 
Figure 8: System function design 
2. Development environment 
In this work, some algorithms are written in the Microsoft 
Visual Studio 2013 environment using C++ language, 
which is efficient and stable. In addition, some functions 
in OpenCV are used for image processing, Caffe's deep 
learning architecture is used for book recognition, and 
MySQL is used for database management. 
(1) Microsoft Visual Studio 2013 
As an integrated solution, Microsoft Visual Studio 2013 
is suitable for all groups, whether it is individuals or 
development teams of different sizes. It is currently the 
most popular integrated development environment for the 
Windows platform. As a development tool, it has fast 
application development, efficient teamwork, and 
innovative user experience. Excellent debugging 
functions are often used for visual programming. 
(2) OpenCV 
OpenCV is an open-source computer vision library, 
which covers the most advanced computer vision and 
cloud machine preparation sub-learning.OpenCV is not 
only compatible with interface languages such as C/C++, 
Python, and Matlab, but also compatible with multiple 
operating systems such as Windows, Linux, and Mac OS. 
Due to its high efficiency, openness, independence, and 
other advantages, it is very suitable for real-time image 
processing, so some functions of the OpenCV3.0 version 
are used in the development process. 
(3) Caffe 
Caffe (Conventional Architecture for Feature Extraction) 
is an expressive, innovative, efficient, and 
easy-to-extensible tool. It is highly appreciated by 
researchers in the design of deep learning network 
frameworks. Caffe focuses on machine vision, adopts 
C++/CUDA architecture, has command line, Python, and 
Matlab interfaces, and has a wealth of open-source 
models, demonstrations, and in-depth learning examples. 

Research on Automatic Recognition Technology of Library Books…                      Informatica 48 (2024) 29–40   37 
The most important feature is the ability to seamlessly 
switch between CPU and GPU. 
(4) MySQL 
MySql is a relational database, which uses the standard 
form of the SQL language. It is now one of the most 
popular relational data management systems. MySQL is 
small and fast. Since the code is open source, it also 
reduces the cost of ownership. The system is written in C 
and C++, creating a portable source code. 
4.2 Detailed design of main functional 
modules 
1. Preprocessing module 
As mentioned above, the process is based on the edge 
analysis and direction-constrained line tracking algorithm 
proposed in this paper. The first algorithm attenuates the 
spine images collected by the image manager; On this 
basis, according to Canny's point of view, the boundary of 
the spine in the edge image is a straight line. Using the 
straight-line tracking algorithm proposed in this paper, 
the long straight-line segment is detected as a candidate, 
and the critical point is selected. As the candidate line 
segment for the last spine, the edge is used for 
segmentation; the initial image is segmented along the 
identified edge line of the vertebrae, and the rectangular 
frame around the book is removed. Because the book is 
tilted, the rectangular frame must be tilted to get the 
image of the spine of the book. 
 
2. Identity Module 
The recognition part only needs to call the network model 
trained through the C++ interface provided by Caffe. It is 
necessary to scale the single-character image sequence 
received by the preprocessing module to a specific size 
(28×28) network that satisfies the input conditions, and 
the average image is reduced (whitened) and then flows 
into the organization for further calculations to obtain the 
recognition result. Check the maximum probability in the 
softmaxlayer, and read the identification number of the 
maximum value, which is the result of input recognition. 
 
3. Post-processing module 
The post-processing module of the system is the process 
of analyzing, classifying, and returning the previously 
obtained verification results. This includes interaction 
with existing library databases. Due to unclearness and 
other reasons, part of the phone number is missing. This 
document uses regular expressions to perform fuzzy 
query processing on data to achieve higher accuracy. The 
specific implementation code is as follows: 
 
for (int i=0; i<= book[j].length () +1; i= i+ book[j].length 
()) { 
str = book[j].insert(1," %"); 
} 
mysql_query(con, "set names gbk"); 
query = "select booked from book where booked like 
"+str+" limit 1"; 
strcpy(q, query.c_str()); 
rt =mysql_real_query (con, q, strlen (q)); 
4.3 Test and analysis of automatic 
identification system of library books 
Test Tools ： In the testing process, to ensure the proper 
operation of the system and to provide a reliable basis for 
the analysis and processing of different data. The first 
step is to select the corresponding functional modules to 
be used. Then the sub-modules are connected as a whole 
according to the actual requirements and tested to verify 
their feasibility before starting the formal operation. This 
experiment mainly adopts two kinds of sensors, namely a 
reed switch and infrared sensor, to simulate the reader 
content and specific location information contained in the 
book image collected from the reader reading related 
information, as well as the identification and processing 
of the library internal staff through different channels in 
the library to the library shelves, paper, etc.; the 
automatic library book identification system based on 
image processing is designed according to different needs. 
Through this algorithm, the information of bookshelves 
and paper in the library is extracted effectively, so that the 
readers can get the satisfaction of their psychological 
feelings according to the requirements in the reading 
process. The experiment also analyzes the specific 
detection of book features and bibliographic contents by 
two sensor models, English RGB and SIFT, based on 
image pre-processing technology recognition. The 
experimental results show that the book recognition 
system based on image processing can effectively detect 
the contents involved in the reading process of readers. 
The algorithm can recognize the book information 
accurately and without errors. 
System Performance ： This paper designs an automatic 
library book recognition system based on image 
processing technology. By comparing various kinds of 
books and journals, it is selected among them which are 
suitable for the research and production use of this 
subject, the best economic applicability, and the 
conditions of the practical situation are generally applied 
to various fields. The algorithm has good performance in 
terms of hardware, and it can meet the laboratory 
real-time detection requirements. This can effectively 
solve the current problem of not being able to provide 
high-performance data transmission channels in a 
large-scale integrated environment; at the same time, the 
highly automated characteristics of image processing 
technology make the system highly flexible and adaptable 
to a variety of different scenarios and can be realized 

38   Informatica 48 (2024) 29–40                                                                       H. Xun 
according to different needs for various types of image 
information and retrieval results, etc., which makes the 
system can be well applied in a variety of application 
scenarios This enables the system to be used in a variety 
of application scenarios. 
 
4.4 Compared with other current 
frameworks 
The strength of our suggested method is its capacity to 
recognize the category of every book in a picture, 
regardless of how the volumes are stacked. Table 1 
compares the current framework in use with other 
frameworks. Our technique Figure 11 achieves accurate 
categorization of book kinds, which goes beyond prior 
methods like the one described in Figure 9, which only 
focused on spine segmentation, and Figure 10, which 
expanded to spine segmentation and text extraction. 
Notably, current techniques were unable to identify the 
particular genre of books. They demonstrate the 
superiority of our suggested strategy over other current 
frameworks for book categorization through a 
comparative examination. 
 
Table 1: Comparison of alternative frameworks currently 
in use 
Methods Book Spine 
Segmentation 
Accuracy 
(%) 
Text 
Extraction 
Accuracy 
(%) 
Book type 
categorization 
Accuracy (%) 
Method 
1 [20] 
92 0 0 
Method 
2 [21] 
95 86 0 
Method 
3 [22] 
95 87 90 
Proposed 97 94 95 
 
Figure 9: Accuracy of Book Spine Segmentation 
. 
Figure 10: Accuracy of text extraction 
 
Figure 11: Accurate classification of book types 
 
5  Discussion 
 
The article would be greatly improved by include a 
thorough examination of the performance of the 
recommended system in a separate discussion section. To 
fully understand the distinctiveness and efficacy of the 
suggested approach, a comparison with analogous efforts 
is required. Gaining insight into the unique contributions 
of the proposed system requires an understanding of any 
observed performance differences. This improvement 
provides more than just a synopsis; it enables readers to 
explore the system's useful implications and applicability 
in a wider range of scenarios. This thorough explanation 
helps readers understand the system's importance and 
possible ramifications by clarifying its advantages and 
disadvantages in relation to current methods. In the end, 
this method provides a nuanced comprehension of the 
contributions made by the proposed system, enabling a 
better assessment of its advantages and disadvantages in 
practical situations. 
 
6  Conclusions 
Active library management uses modern management 
methods to provide borrowing and intelligent 
management, to adapt to changing times. Based on the 
technical status of the library, this research explores and 
analyzes the application of automatic identification 
technology in disordered books, and designs a specific 
system. Only the correct level can be guaranteed, the 
straight lines in the image can be developed faster, and 

Research on Automatic Recognition Technology of Library Books…                      Informatica 48 (2024) 29–40   39 
the noise reduction can be improved. The results of 
current experiments make this algorithm suitable for 
machine vision and image processing applications in 
terms of time, efficiency, and reliability. To build a fully 
automatic identification system, the construction and 
design of each link in the system must be completed. The 
modules in the currently designed automatic 
identification system include pre-processing, 
identification, and post-processing modules. Each module 
plays a different role in the entire system, but it is 
indispensable. This thesis researches the automatic library 
book recognition technology based on image processing, 
mainly by classifying different book labels to achieve the 
reader input reading mark on the shelf, and then get the 
desired information according to the corresponding 
characteristics. Through experiments and practical tests, 
certain results have been achieved. The experimental 
results show that the automatic book recognition system 
based on image processing is effective and efficient in 
label classification and information retrieval. Some of the 
limitations include the inability to segment incomplete 
images using conventional threshold approaches, the 
noise sensitivity of some edge detection algorithms, and 
possible inefficiencies in the fixed-threshold 
implementation of the Canny edge detection algorithm. 
Because it depends on particular sensors, the book 
recognition system could not be as flexible in different 
library settings. 
 
Data availability 
The data used to support the findings of this study are 
available from the corresponding author upon request. 
 
Conflicts of Interest 
The authors declare no conflicts of interest 
Funding Statement 
This study did not receive any funding in any form. 
References 
[1] Senkyire, I.B. and Liu, Z., 2021. Supervised and 
semi-supervised methods for abdominal organ 
segmentation: A review. International Journal of 
Automation and Computing, 18(6), pp.887-914.. 
[2] Yi, P.A.N., Jin, L.I.U., Xu, T.I.A.N., Wei, L.A.N. 
and Rui, G.U.O., 2021. Hippocampal segmentation 
in brain mri images using machine learning 
methods: A survey. Chinese Journal of 
Electronics, 30(5), pp.793-814.. 
[3] Fan, H., Sun, Y., Zhang, X., Zhang, C., Li, X. and 
Wang, Y., 2021. Magnetic-resonance image 
segmentation based on improved variable weight 
multi-resolution Markov random field in 
undecimated complex wavelet domain. Chinese 
Physics B, 30(7), p.078703.. 
[4] Yi, P.A.N., Jin, L.I.U., Xu, T.I.A.N., Wei, L.A.N. 
and Rui, G.U.O., 2021. Hippocampal segmentation 
in brain mri images using machine learning 
methods: A survey. Chinese Journal of 
Electronics, 30(5), pp.793-814.. 
[5] Zhang, Y. and Tian, Y., 2021. A new image 
segmentation method based on 
fractional-varying-order differential. Journal of 
Beijing Institute of Technology, 30(3), pp.254-264.. 
[6] Li, Y., Blois, G., Kazemifar, F. and Christensen, 
K.T., 2021. A particle-based image segmentation 
method for phase separation and interface detection 
in PIV images of immiscible multiphase 
flow. Measurement Science and Technology, 32(9), 
p.095208.. 
[7] Yongbin, Y.U., Chenyu, Y.A.N.G., Quanxin, 
D.E.N.G., Tashi, N., Shouyi, L. and Chen, Z., 2021. 
Memristive network-based genetic algorithm and its 
application to image edge detection. Journal of 
Systems Engineering and Electronics, 32(5), 
pp.1062-1070.. 
[8] Liu, X. and Richardson, A.G., 2021. Edge deep 
learning for neural implants: a case study of seizure 
detection and prediction. Journal of Neural 
Engineering, 18(4), p.046034.. 
[9] Gong, Y., Liu, Y. and Yin, C., 2021. A novel 
two-phase cycle algorithm for effective cyber 
intrusion detection in edge computing. EURASIP 
Journal on Wireless Communications and 
Networking, 2021(1), pp.1-22.. 
[10] Xu, Z., Ji, X., Wang, M. and Sun, X., 2021, June. 
Edge detection algorithm of medical image based on 
Canny operator. In Journal of Physics: Conference 
Series (Vol. 1955, No. 1, p. 012080). IOP 
Publishing.. 
[11] Yu, T., 2021, February. Computer Application of 
Edge Detection Of Substrate Image Considering 
Fuzzy Microflora Optimization Algorithm. 
In Journal of Physics: Conference Series (Vol. 
1792, No. 1, p. 012033). IOP Publishing.. 
[12] Gustavo Z. Felipe, Jacqueline N. Zanoni, Camila C. 
Sehaber-Sierakowski, Gleison D. P. Bossolani, Sara 
R. G. Souza, Franklin C. Flores, Luiz E. S. Oliveira, 
Rodolfo M. Pereira,Yandre M. G. Costa. Automatic 
chronic degenerative diseases identification using 
enteric nervous system images[J]. Neural 
Computing and Applications, 2021, 33(22):310-328. 
 

40   Informatica 48 (2024) 29–40                                                                       H. Xun 
[13] Bhardwaj, C., Jain, S. and Sood, M., 2021. Transfer 
learning based robust automatic detection system for 
diabetic retinopathy grading. Neural Computing and 
Applications, 33(20), pp.13999-14019.. 
[14] Niu, Y., Fang, L., Sun, S., Jiang, J. and Chen, P., 
2018. The Design of Book Sorter Base on Radio 
Frequency Identification. Journal of Applied 
Science and Engineering Innovation, 5(1), 
pp.18-21.. 
[15] Ostler, K.R., Norlander, B. and Weber, N., 2021. 
Using open data to inform public library branch 
services. Public Library Quarterly, 40(4), 
pp.365-377.. 
[16] Wang, X., Li, B. and Geng, Q., 2012, August. 
Runway detection and tracking for unmanned aerial 
vehicle based on an improved canny edge detection 
algorithm. In 2012 4th International Conference on 
Intelligent Human-Machine Systems and 
Cybernetics (Vol. 2, pp. 149-152). IEEE.. 
[17] Chen, Q., Xiao, L. and Zhuang, S., 2012, May. A 
New Data Reduction Approach over the Stream 
Processor Architecture. In 2012 IEEE 26th 
International Parallel and Distributed Processing 
Symposium Workshops & PhD Forum (pp. 
2300-2304). IEEE.. 
[18] Li, R., Wu, H., Liu, S., Rahman, M.A., Liu, S. and 
Kwok, N.M., 2018, April. Image edge tracking via 
ant colony optimization. In Ninth International 
Conference on Graphic and Image Processing 
(ICGIP 2017) (Vol. 10615, pp. 569-576). SPIE.. 
[19] Lu, J., Cai, Z., Yao, B. and Chen, B., 2020. A novel 
engagement-pixel image edge tracking method for 
extracting gear tooth profile edge. Proceedings of 
the Institution of Mechanical Engineers, Part C: 
Journal of Mechanical Engineering Science, 234(2), 
pp.405-416.): 
[20] Jubair, M.I. and Banik, P., 2013, July. A technique 
to detect books from library bookshelf image. 
In 2013 IEEE 9th International Conference on 
Computational Cybernetics (ICCC) (pp. 359-363).  
[21] Tabassum, N., Chowdhury, S., Hossen, M.K. and 
Mondal, S.U., 2017, February. An approach to 
recognize book title from multi-cell bookshelf 
images. In 2017 IEEE International Conference on 
Imaging, Vision & Pattern Recognition 
(icIVPR) (pp. 1-6).  
[22] Fatema, K., Ahmed, M.R. and Arefin, M.S., 2022. 
Developing a system for automatic detection of 
books. In Second International Conference on 
Image Processing and Capsule Networks: ICIPCN 
2021 2 (pp. 309-321). Springer International 
Publishing.