Realtime Detection of Table Objects and Text Areas for OCR
Preprocessing
HANGSEO CHOI, JONGPIL JEONG,
Department of Smart Factory Convergence,
Sungkyunkwan University,
Cheoncheon-dong, Jangan-gu Suwon-si, Gyeonggi-do,
REPUBLIC OF KOREA
Abstract: - OCR (Optical Character Recognition) is a technology that automatically detects, recognizes, and
digitally converts text into images. OCR has a variety of uses, including reducing human error when viewing
and typing documents and helping people work more efficiently with documents. It can increase efficiency and
save money by eliminating the need to manually type text, especially when scanning documents or digitizing
images. OCR is divided into text object detection and text recognition in an image, and preprocessing
techniques are used during the original document imaging process to increase the accuracy of OCR results.
There are various preprocessing techniques. They are generally classified into image enhancement, binarization
techniques, text alignment and correction, and segmentation techniques. In this paper, we propose a special-
purpose preprocessing technique and application called Table Area Detection. Recently, table detection using
deep learning has been actively researched, and the research results are helping to improve the performance of
table recognition technology. Table detection will become an important preprocessing technology for text
extraction and analysis in various documents, and it requires a lot of research and accuracy. While many
previous studies have focused on improving the accuracy of OCR algorithms through various techniques, this
study proposes a method to discover and exclude false positives by introducing a factor called Table Area
Detection.
Key-Words: - Text Detection, Text Recognition, Table Detection, Object Detection, Preprocessing,
Segmentation
Received: June 17, 2022. Revised: April 22, 2023. Accepted: May 21, 2023. Published: June 21, 2023.
1 Introduction
OCR is divided into two phases: text detection and
text recognition. The text detection stage identifies
and extracts the regions with text from the image,
while the text recognition stage recognizes the
actual text from the detected text regions and
converts it into computer-understandable text. These
two stages are closely linked to improving the
performance of the OCR system, and preprocessing
is located at the initial stage of text detection.
The purpose and method of preprocessing is to
eventually provide the optimal conditions for
detecting and recognizing text in an image, but
image quality is not the only thing that can get in the
way of identifying text. Many papers and
experiments have focused on improving OCR
performance by preprocessing documents, but the
characteristics of documents vary so much that
preprocessing is not the answer. A typical example
is table detection. To detect text in a document, it is
important to correctly identify the text area, and
tables complicate the text structure. Between the
rows and columns of a table, there are various lines,
grids, borders, etc. that create confusion in the
process of identifying text areas. Text detection
algorithms use a variety of computer vision
techniques to detect boundaries, segment text
regions, and recognize text blocks. They analyze the
brightness, color, texture, and shape of the image to
determine the features of the text and identify
regions. Tables need to be separated from text at this
stage due to the difficulty in detecting boundaries,
text overlap, and interference.
Table detection is a task closely related to text
detection and recognition in OCR, [1]. Tables are a
form of structured text, consisting of rows and
columns, with each cell containing textual
information. Therefore, it is necessary to understand
the structure of the table and extract the contents of
each cell, which involves detecting row and column
boundaries, identifying regions in each cell, and
recognizing text in each cell, [2]. The recognition
rate of OCR depends on the quality of the image,
[3], and rather than finding the location and size of
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.23
Hangseo Choi, Jongpil Jeong
E-ISSN: 2224-3402
197
Volume 20, 2023
the text in the image, it extracts the text information
from the entire area within the image and then
processes it. However, targeting OCR based on the
table area rather than the entire document, and
excluding the text area from recognition if it
exceeds the table area, will improve the accuracy of
the detection target and reduce boundary
interference, [4]. This proposal locates the table in
the image, extracts the position and size of each cell,
and text information, and excludes areas where it is
technically difficult to separate the text. Table
recognition techniques typically use object detection
techniques to locate tables. Object detection is a
technique for finding the location and size of objects
in an image, [5]. It compares the respective
coordinates of the table and text, measures the
distance between the two coordinates, and sets a
certain threshold value to exclude the text area from
the OCR target when it exceeds the area of the table
object, [6].
The algorithm for detecting and separating table
and text regions is not different from the technique
for extracting text information from images.
Therefore, the text recognition rate can be improved
by clearly specifying the extraction target. In OCR
preprocessing, table positions, and text areas can be
detected, and table and text boundaries can be
separated to improve recognition rates in the text
recognition stage. Alignment and correction of text
areas can also be performed. By accurately aligning
the rows and columns of the table and adjusting the
regular placement of the document, text recognition
accuracy can be improved.
The paper is organized as follows: Section 2
provides an overview of the technology and the
concept of OCR using deep learning, Section 3
presents the overall architecture of the system,
Section 4 describes the implementation process, and
Section 5 concludes with future research
considerations.
2 Related Work
2.1 Faster R-CNN
Faster R-CNN(Faster Regions with Convolutional
Neuron Networks features) is an algorithm proposed
in 2015, [7], that can perform fast and accurate
object detection by compensating for the
shortcomings of R-CNN, [8], and Fast R-CNN. It
first processes the input image with a
CNN(Convolutional Neural Network) to generate a
feature map. It then uses an RPN(Region Proposal
Network) to generate candidate regions and
performs RoI(Region of Interest) pooling on these
candidate regions to extract the features of each
object. These features are then utilized to perform
object classification and bounding box estimation.
Recently, there has been a lot of research in the field
of table recognition that utilizes it to detect table
regions. By utilizing it, table areas can be detected
accurately, and tables of various sizes and shapes
can be detected.
2.2 YOLO
YOLO (You Only Look Once) is an algorithm
proposed in 2016 that provides fast speed and high
accuracy. It divides the input image into a grid and
predicts the probability of the bounding box and
corresponding object in each grid cell. It uses these
predictions to perform object classification and
bounding box estimation. It has been widely used in
the field of table recognition recently due to its fast
speed and high accuracy. YOLOv3, [9], provides
both high accuracy and fast speed, and it can detect
tables of various sizes and shapes. Recently, various
object detection algorithms, including Faster R-
CNN and YOLO, have been developed and continue
to be used in the field of table recognition to provide
high recognition rates and fast speeds.
2.3 Mask R-CNN
Mask R-CNN, [10], is a state-of-the-art object
detection and instance segmentation algorithm that
has been proven to be effective for a variety of
tasks, including table segmentation. It is an
extension of Faster R-CNN, which performs object
detection and object segmentation simultaneously,
[11]. Therefore, a method is proposed to utilize it to
detect table regions within document images and
perform table recognition based on them. This
method is performed in the following steps.
Step 1: Perform object detection and object
segmentation in the image.
Step 2: Extract table regions from the object
segmentation results.
Step 3: Perform table recognition based on the
extracted table regions, [11].
In recent years, there have been many advances
in it that have improved its performance on table
segmentation tasks. One of the most important
advances is the use of attention mechanisms, which
allow the model to focus on specific regions of the
image when making predictions. They are
particularly effective for segmenting tables, as
tables are often difficult to distinguish from other
objects in an image. Another important development
is the use of data augmentation techniques. Data
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.23
Hangseo Choi, Jongpil Jeong
E-ISSN: 2224-3402
198
Volume 20, 2023
augmentation techniques are used to artificially
increase the size of a training dataset. This can help
improve the model's performance on unseen data.
Many improvements have been made to its
architecture of it, which can be used to segment
tables in a variety of documents, including scientific
papers, news articles, and legal documents.
2.4 Table Net
End-to-end models that perform table recognition
and OCR simultaneously are a relatively new area
of research, [12]. Deep learning techniques and
large datasets for table recognition, [13], and OCR
have become available, and because of these
advances, end-to-end models, [14], that perform
table recognition and OCR simultaneously are
becoming increasingly common. These models have
several advantages over traditional approaches.
First, they are more accurate. Traditional approaches
often perform table recognition and OCR separately.
This can lead to errors because the results of one
step can affect the results of the other. End-to-end
models, on the other hand, perform table recognition
and OCR simultaneously, [15], which helps ensure
the accuracy of the results. Second, it's more
efficient. Traditional methods for table recognition
and OCR can be time-consuming. The end-to-end
model, on the other hand, is much faster. This is
because it performs both tasks simultaneously.
Third, it's more flexible. Traditional table
recognition and OCR methods are often limited to
certain types of tables, while end-to-end models can
be used to recognize, [16], and OCR a wide variety
of tables. Overall, end-to-end models that perform
table recognition and OCR simultaneously are
promising new approaches. They are more accurate,
efficient, and flexible than traditional approaches, so
they are likely to become increasingly common in
the future.
2.5 OCR
A typical example is deep learning-based OCR.
Deep learning algorithms are used to perform
character recognition, which creates a model to
classify characters by learning features extracted
from the character area. Typically, a convolutional
neural network (CNN) is used. CNNs are
responsible for extracting features within an image
and recognizing characters based on the extracted
features. The second is template matching-based
OCR. Template matching recognizes characters by
calculating the similarity between the input image
and the template image. A template image for each
character is created in advance, and the most similar
character is recognized by measuring the similarity
between the input image and the template image.
However, this method is vulnerable to variations in
size, rotation, and distortion, and requires the
preparation of many template images. Finally,
statistical-based OCR recognizes characters by
analyzing their statistical characteristics. A
statistical model learns the frequency, occurrence
pattern, probability distribution, etc. of each
character, and recognizes the characters in the input
image based on the statistical model. This method
makes good use of the statistical characteristics of
language to improve recognition performance, but it
requires a large amount of training data and requires
the use of models specialized for a particular
language.
3 Proposed Method
3.1 System Process Flow
The main process performs preprocessing on the
input image, as shown in Fig. 1, and then separates
the table and text areas. If there is text outside the
table area or overlapping, it is processed separately,
and text recognition is performed on the text that
exists in the area.
Fig. 1: System Process Flow
After image input and preprocessing, the
process is divided into two stages: Table Detection
and Text Detection. Each step extracts only table
and text information while keeping the document
structure intact. In step 1, Faster R-CNN was used
to train table detection.
Faster R-CNN combines two main components,
RPN and Fast R-CNN, to effectively detect objects
and generate bounding boxes. Since table detection
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.23
Hangseo Choi, Jongpil Jeong
E-ISSN: 2224-3402
199
Volume 20, 2023
involves identifying rectangular regions of a certain
shape within an image or document, we believe that
Faster R-CNN, which simultaneously performs
object detection and bounding box generation, is
suitable for table detection. Fig. 2 shows the
structure of Table Detection, including Faster R-
CNN.
Fig. 2: 1st SETP Table Detection
In step 2, we used CRAFT (Character Region
Awareness for Text Detection) to detect text, as
shown in Fig. 3. CRAFT is a deep learning-based
algorithm for character region detection that
specializes in detecting character regions in images
containing text, and can accurately identify the
location and shape of the text.
Fig. 3: 2nd STEP Text Detection
The idea of this paper consists of finding a table
in an image and extracting the position, size, and
text information of each cell, and the main steps are
shown below.
Table recognition with object detection, [17]:
Use the image dataset and the table bounding
box information of the image to find the table
in the image using Faster R-CNN and train the
object detection model.
Segment the table region: segment the table
region in the original image using the table
bounding box information obtained from the
object detection model. The segmented table
image is preprocessed for processing in the
next step.
Cell segmentation: segment the cell region by
finding the structure of rows and columns in
the table image. For this purpose, we used
Hough transform, a line detection algorithm,
[18], but finally used an object detection
model.
Text extraction via OCR: We use the easy
OCR engine to extract the text of each cell
region. Preprocess the extracted text to remove
noise and extract coordinates for the text area.
Filter text area: Compares the area information
of the table object with the text coordinates of
each cell and excludes the text area from OCR
if it exceeds the area of the table object based
on the threshold value you set.
3.2 Table Detection
To use the Faster R-CNN model, we first define the
model architecture, which uses a backbone network
based on ResNet-50 and an FPN(Feature Pyramid
Network) to extract feature maps. ResNet-50 is a
50-layer residual network, a deep CNN with
excellent performance in object recognition and
classification, which we used as the basic structure
to extract features from images. The FPN detects
objects of different sizes by extracting feature maps
of different sizes, and it takes the feature maps
extracted by ResNet-50 as input and incorporates
features from higher levels into lower levels to
produce an improved feature representation.
Faster R-CNN consists of RPN, which suggests
RoI in an image, and Fast R-CNN, which classifies
these regions and refines the bounding box. Model
training is prepared by configuring RPN.
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.23
Hangseo Choi, Jongpil Jeong
E-ISSN: 2224-3402
200
Volume 20, 2023
Fig. 4: Towards Real-Time Object Detection with
RPNs, [7]
Fig. 4 shows the structure of an RPN that
generates bounding box candidates for an object
while sliding over an input image or feature map
using anchor boxes of different sizes and
proportions.
The RPNs are used to generate anchor boxes for
bounding box regression and classification. The
anchor box generates anchor boxes at each location
in the feature map. Anchor boxes are reference
Bounding Boxes with different sizes and ratios to
correspond to different object sizes and aspect ratios
in the image. Multiple anchor boxes are generated at
each location with anchor sizes and ratios defined
by hyperparameters.
RPNs are trained using a multi-task loss
function consisting of a binary classification loss
and a bounding box regression loss. The binary
classification loss typically uses a binary cross-
entropy loss, and the regression loss typically uses a
smooth L1 loss. The two losses are combined to
compute the final loss, and the weights are updated
to minimize it. The loss function formula for this
process is as follows. Binary Cross-Entropy Loss
The binary cross-entropy loss is used by the RPN to
classify whether each anchor box is an object or
background. The formula is as follows.
: Binary classification loss
: Mini batch size
: actual label of the i-th anchor box (object: 1,
background: 0)
: Model's predicted probability for the i-th
anchor box (probability of being an object)
Bounding Box Regression Loss (Smooth L1
Loss) The Bounding Box Regression Loss is used in
the process of aligning the coordinates of the anchor
boxes in the RPN with the actual bounding boxes.
The formula is as follows.
: Bounding Box regression loss
: Mini batch size
: Actual Bounding Box coordinates of the i-th
anchor box (x, y, w, h)
: Model's predicted Bounding Box coordinates
for the i-th anchor box (x, y, w, h)
SMOOTH_L1: Smooth L1 loss function
Now combine the two losses to calculate the final
loss.
: Final loss
λ: Weight of the Bounding Box regression loss
(set as a hyperparameter)
Update the weights of the RPN model in the
direction of minimizing the final loss L_total
calculated in this way. Once the object detection
results for the image are obtained, filter out the
bounding boxes with low confidence among the
model inference results, or apply NMS(Non-
Maximum Suppression) to remove redundant
bounding boxes. From these filtered and refined
results, we extract the corresponding bounding box
coordinates in a table and prepare them for
comparison with the text extraction area.
3.3 Text Detection
CRAFT performs the steps of extracting the features
of an image to detect text regions. It mainly uses
CNN to generate feature maps at different scales,
which refers to the process of extracting structural
information about text to identify text regions in an
input image. CNN learn regional patterns in an
image and uses them to distinguish between text and
non-text regions in an image. Passing an image
through multiple CNN layers creates a multi-level
feature map, which can capture different sizes and
details of text areas. A small-scale feature map helps
recognize small text areas, while a large-scale
feature map helps recognize large text areas. The
generated text candidate regions are then adjusted
with a bounding box in a subsequent step. Feature
maps are also used to extract text structure
information. Feature maps are good at capturing text
features such as sharp boundaries, strong vertical
lines, text orientation, etc., and utilize this
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.23
Hangseo Choi, Jongpil Jeong
E-ISSN: 2224-3402
201
Volume 20, 2023
information to accurately detect text regions. Fig. 6
shows an example of text detection from a table in a
document.
Fig. 5: Text Detection by CRAFT
3.4 Text Recognition
Text Recognition uses the easyOCR algorithm.
easyOCR is a deep learning-based character
recognition technique that aims to recognize the
bounding box of a text and follows the following
steps.
Preprocessing: The preprocessing step is
important to improve the quality of the image
and make it easier for the deep-learning model
to classify the characters. Common
preprocessing techniques include removing
noise from the image, such as dust, scratches,
and uneven lighting, and adjusting the contrast
of the image to make the characters more
visible.
Segmentation: The segmentation step is used to
break down the image into individual
characters. This is done by identifying
connected components in the image. Connected
components are groups of pixels that are all
connected to each other. Connected
components that are the same size as a
character are then categorized into characters,
[19].
Classification: The classification stage is used
to identify individual characters in an image,
[20] This is done using a deep learning model.
The deep learning model is trained on a large
dataset of images containing labeled characters.
The model learns to identify characters by their
shape, size, and texture.
Reconstruction: The reconstruction step is used
to combine individual characters into text, [21].
This is done by aligning the characters
according to their bounding boxes and then
combining them together.
4 Experiment and Results
4.1 Experimental Environments
This experiment was run on Google Corel Pro with
the following hardware: GPU NVIDIA T4, CPU
Inter(R) Xeon(R) CPU, 16GB RAM Software:
Development language: Python 3.8, CUDA version:
11.8, Deep Learning framework: PyTorch 1.8.1,
Required Python libraries: OpenCV, NumPy, PIL,
etc. To build the environment by installing the
necessary Python packages and libraries, we
customized the SR (Shipping Request) document,
which is used in the field for test data other than
training. Colab is an online implementation of
Jupiter Notebook that allows you to write and run
Python code, and you can use PyTorch to
implement the table detection model and text
recognition model.
4.2 Data Set
The model was trained using a prepared custom
dataset and the Faster R-CNN model. The custom
dataset utilized delivery request documents used in
the field. The customized documents were collected
from actual documents used in the shipping and
logistics industry and consisted of 4 templates from
27 companies, totaling 112 files. Each file contains
titles, diagrams, text, and images such as company
logos. This data was used for testing purposes rather
than training purposes and was not pre-trained.
4.3 Table Detection by Faster R-CNN
It uses the Faster R-CNN to find a table in an image.
The document image was resized and normalized to
preprocess it into a format suitable for the Faster R-
CNN model. Although Faster R-CNN does not
require a fixed input size, we resized the image to a
reasonable size of 800-1000 pixels wide by 1000
pixels high, considering GPU memory limitations
and training time, model performance, and
computational efficiency, while maintaining the
aspect ratio of the original image. The reason for
normalizing the image is to stabilize the training
process by making the distribution of pixel values
constant and to speed up convergence, so we scaled
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.23
Hangseo Choi, Jongpil Jeong
E-ISSN: 2224-3402
202
Volume 20, 2023
the image pixel values to the range [0, 1], and then
performed normalization. As shown in Fig. 5, you
can build a single pipeline to extract surface areas
and table areas in real-time.
Fig. 6: Detect Table Process and results
4.4 Text Detection by CRAFT
The experiments were divided into 1 and 2:
Experiment 1 recognized only text without table
detection, and Experiment 2 tried to recognize text
in the table area with table detection results. Since
CRAFT is a pre-trained model for text detection, we
were able to extract text directly from the image
without having to train a separate object detection
model. However, when tables in the document
overlapped with table areas, it sometimes failed to
recognize the table structure. The Text Detection
Results are presented in Fig. 7.
Fig. 7: Text Detection Results
4.5 Results
For Experiment 1, we performed text extraction
using only CRAFT and easyOCR to check the text
detection rate without table detection. To perform
Experiment 2, we prepared a custom dataset to train
the Faster R-CNN model, extracted the bounding
box to obtain the coordinate information, and then
segmented the table area with the table recognition
output coordinates and extracted the text
information with Faster R-CNN to obtain the
coordinates of each text area. To compare the
coordinates of the most important table and text
regions during the implementation phase, we
measured the distance between the two coordinates
and set a certain threshold to exclude the text region
from OCR if it exceeds the region of the table
object. This allowed us to extract only the text of the
table and check the text extraction accuracy
compared to the steps in Experiment 1, and we
found a performance improvement of about 3%. The
results in Table 1 show that there are a total of 4
document template types, and the more templates
you train, the higher the table detection rate.
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.23
Hangseo Choi, Jongpil Jeong
E-ISSN: 2224-3402
203
Volume 20, 2023
Table 1. Table detection rate as training
documentation templates increase
Template
Data
Table Detect rate
4
112
82%
3
112
73%
2
112
70%
1
112
65%
The results in Table. 2 show the results of
Experiments 1 and 2 separately. Experiment 1
(column name: Text Detect 1) only tried to detect
text without detecting tables. Experiment 2 (column
name: Text Detect 2) detects the table and text,
removes duplicates, and detects the text. As a result,
we can see that removing and detecting duplicates in
the table can contribute to performance
improvement.
Table. 2. Plain text detection rate and table detection
rate after excluding duplicate regions
Table Detect rate
Text Detect 1
Text Detect 2
82%
86%
89%
5 Conclusions
In this paper, we presented an approach to detecting
tables in images, extracting the location, size, and
textual information of each cell, and comparing
them to text regions. The efficient table recognition
and text extraction pipeline is a fusion of deep
learning models and commercial open source. The
pipeline includes all the necessary steps to process
information related to tables in document images,
and we believed that a method that measures the
distance between two coordinates and excludes
them from OCR if the text area exceeds the area of
the table object according to a certain threshold
would improve the text detection rate compared to
recognizing an unspecified number of documents. In
addition, a coordinate-based region overlap filtering
method to implement this concept would help
improve the accuracy of text extraction and preserve
the structure of the table. Nevertheless, we were
disappointed that we had to use a custom dataset
due to the lack of data. As it may be difficult to
ensure the generalization of the model in various
situations, we expect that repeating the same
training process and experiments with an advanced
model trained on a large dataset would lead to
improved performance and universality.
Acknowledgment:
This research was supported by the SungKyunKwan
University and the BK21 FOUR(Graduate School
Innovation) funded by the Ministry of
Education(MOE, Korea) and the National Research
Foundation of Korea(NRF). And this work was
supported by the National Research Foundation of
Korea (NRF) grant funded by the Korea government
(MSIT) (No. 2021R1F1A1060054). Corresponding
authors: Professor Jongpil Jeong.
References:
[1] M. Kasem, A. Abdallah, A. Berendeyev, E.
Elkady, M. Abdalla, M. Mahmoud, M.
Hamada, D. Nurseitov, I. Taj-Eddin, Deep
learning for table detection and structure
recognition: A survey, arXiv:2211.08469v1
[cs.CV], 2022.
[2] Y. Li, L. Gao, Z. Tang, Q. Yan, Y. Huang, A
GAN-Based Feature Generator for Table
Detection, IEEE Transactions on Pattern
Analysis and Machine Intelligence, 03
February 2020.
[3] D. G Lee. CNN-based Image Rotation
Correction Algorithm to Improve Image
Recognition Rate, The Journal of The Institute
of Internet, Broadcasting and Communication
(IIBC) Vol. 20, No. 1, 2022, pp.225-229,
JIIBC 2020-1-32.
[4] C. B. Jang, Implementation of Pre-Post
Process for Accuracy Improvement of OCR
Recognition Engine Based on Deep-Learning
Technology, Journal of Convergence for
Information Technology, Vol. 12. No. 1, 2022,
pp. 163-170.
[5] J. S Choi, Table Detection Scheme based on
Deep Learning, Proceedings of the Korean
Computer Conference, 2022, 930 – 932.
[6] J. Hu, R. S. Kashi, D. Lopresti, and G. T.
Wilfong, Evaluating the performance of table
processing algorithms, International Journal
on Document Analysis and Recognition, Vol.
4, 2022, 140–153,
[7] S. Ren, K. He, R. Girshick, J. Sun, Faster R-
CNN: Towards 14 real-time object detection
with region proposal networks, in Neural
Information Processing Systems (NIPS), 2015,
pp 91-99.
[8] R. Girshick, J. Donahue, T. Darrell, J. Malik.
Rich feature hierarchies for accurate object
detection and semantic segmentation. IEEE
Conference on Computer Vision and Pattern
Recognition (CVPR), 2014, 580–587.
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.23
Hangseo Choi, Jongpil Jeong
E-ISSN: 2224-3402
204
Volume 20, 2023
[9] J. Redmon, S. Divvala, R. Girshick, A.
Farhadi, You Only Look Once: Unified, Real-
Time Object Detection, e IEEE Conference on
Computer Vision and Pattern Recognition
(CVPR), 2017, pp 7263-7271
[10] K. He, G. Gkioxari, P. Dollar, R. Gir-shick.
Mask R-CNN. IEEE International Conference
on Computer Vision (ICCV), 2017, 2961–
2969
[11] Y.W. Lee, J.Y. Park, CenterMask: Real-time
anchor-free instance segmentation, IEEE
Conference on Computer Vision and Pattern
Recognition (CVPR), 2020.
[12] D. W. Embley, M. Hurst, D. Lopresti, and G.
Nagy, Table-processing paradigms: a research
survey, International Journal of Document
Analysis and Recognition (IJDAR), Vol. 8,
No. 2-3, 2016, pp. 66–86.
[13] I. Kavasidis, S. Palazzo, C. Spampinato, C.
Pino, D. Giordano, D. Giuffrida, and P.
Messina, A saliency-based convolutional
neural network for table and chart detection in
digitized documents, arXiv preprint
arXiv:1804.06236, 2018.
[14] S. Appalaraju, Jasani, B. Kota, B.U, X. Y.
Manmatha, R. Docformer, End-to-end
transformer for document understanding. In:
Proceedings of the IEEE/CVF International
Conference on Computer Vision (ICCV),
2021, pp. 993–1003
[15] Y. Baek, B. Lee, D. Han, S. Yun, H. Lee,
Character region awareness for text detection.
IEEE/CVF Conference on Computer Vision
and Pattern Recognition (CVPR), 2019. pp.
9357–9366
[16] A.W. Harley, A. Ufkes, K.G Derpanis,
Evaluation of deep convolutional nets for
document image classification and retrieval,
13th International Conference on Document
Analysis and Recognition (ICDAR), 2015, pp.
991–995
[17] M. Fan, D.S. Kim, Table region detection on
large-scale PDF files without labeled data.
CoRR, abs/1506.08891, 2015.
[18] B. Gatos, D. Danatsas, I. Pratikakis, and S.J.
Perantonis, Automatic table detection in
document images. In Proc. of ICAPR (2005) -
Volume Part I, ICAPR, 05, Berlin,
Heidelberg. Springer-Verlag, pages 609–618.
[19] S. A. Oliveira, B. Seguin, F. Kaplan, segment:
A generic deep-learning approach for
document segmentation. ICFHR 2018.
[20] R. Child, S. Gray, A. Radford, I. Sutskever,
Generating long sequences with sparse
transformers. arXiv preprint
arXiv:1904.10509, 2019.
[21] D. Deng, H. Liu, X. Li, D. Cai. Pixellink:
Detecting scene text via instance
segmentation. In AAAI. 2018.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
-Hangseo Choi set the research topic and goals,
developed the software, conducted the experiments,
validated, and wrote the paper.
-Professor Jongpil Jeong conceptualized the idea,
presented the methodology, and conducted the
review.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
This research was supported by the SungKyunKwan
University and the BK21 FOUR(Graduate School
Innovation) funded by the Ministry of
Education(MOE, Korea) and the National Research
Foundation of Korea(NRF). And this work was
supported by the National Research Foundation of
Korea (NRF) grant funded by the Korean
government (MSIT) (No. 2021R1F1A1060054).
Corresponding authors: Professor Jongpil Jeong.
Conflict of Interest
The authors have no conflict of interest to declare.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONS
DOI: 10.37394/23209.2023.20.23
Hangseo Choi, Jongpil Jeong
E-ISSN: 2224-3402
205
Volume 20, 2023
