Table-structure Recognition Method

Consisting of Plural Neural Network Modules

Hiroyuki Aoyagi

, Teruhito Kanazawa

, Atsuhiro Takasu

, Fumito Uwano

and Manabu Ohta

Okayama University, Okayama, Japan

National Institute of Informatics, Tokyo, Japan

Keywords:

Table-structure Recognition, Neural Network, PDF, XML.

Abstract:

In academic papers, tables are often used to summarize experimental results. However, graphs are more

suitable than tables for grasping many experimental results at a glance because of the high visibility. Therefore,

automatic graph generation from a table has been studied. Because the structure and style of a table vary

depending on the authors, this paper proposes a table-structure recognition method using plural neural network

(NN) modules. The proposed method consists of four NN modules: two of them merge detected tokens in a

table, one estimates implicit ruled lines that are necessary to separate cells but undrawn, and the last estimates

cells by merging the tokens. We demonstrated the effectiveness of the proposed method by experiments using

the ICDAR 2013 table competition dataset. Consequently, the proposed method achieved an F-measure of

0.972, outperforming those of our earlier work (Ohta et al., 2021) by 1.7 percentage points and of the top-

ranked participant in that competition by 2.6 percentage points.

1 INTRODUCTION

The spread of academic paper databases such as

Google Scholar

, DBLP

and CiNii

has allowed

us to collect papers more easily. In academic pa-

pers, tables are often used to show some data, statis-

tics, and experimental results. Table recognition is

an important issue not only to generate graphs au-

tomatically but also to search and compare such ex-

perimental results. Therefore, Ohta et al. pro-

posed a cell-detection-based table-structure recogni-

tion method (Ohta et al., 2019), and they improved

it introducing neural networks (NNs) (Ohta et al.,

2021).

The method proposed in (Ohta et al., 2021) intro-

duced two NN modules: one estimates implicit ruled

lines (IRLs) and the other generates cells by merg-

ing tokens in a table. They analyzed 156 tables in

the ICDAR 2013 table-structure competition dataset

obel et al., 2013) for which the resultant recall, pre-

cision, and F-measure values for measuring adjacency

relation between cells were 0.951, 0.960, and 0.955,

respectively. These results marginally outperformed

https://scholar.google.com

https://dblp.org

https://ci.nii.ac.jp

those of the top-ranked participant Nurminen (Nur-

minen, 2013) in that competition. The method, how-

ever, had some room for improvement in its process-

ing ﬂow and NN modules.

In this paper, we propose an improved method

of (Ohta et al., 2021) by introducing NN-based hor-

izontal and vertical token mergers and by sophisti-

cating input features for the NN modules. The pro-

posed method also has two NN modules for IRL es-

timation and cell generation. In addition, it merges

horizontally adjacent tokens before estimating IRLs

and merges vertically adjacent tokens after the esti-

mation. We demonstrate the effectiveness of the pro-

posed method by experiments using the ICDAR 2013

table competition dataset. Note that the proposed

method analyzes the structure of born-digital tables

in PDF documents and not of table images because of

the high availability of recent born-digital tables on

the Internet. Also note that the method currently does

not detect tables in PDF documents and we manually

locate them.

This paper is structured as follows. We introduce

related work on table-structure recognition in Section

2 and explain our proposed method and its four NN

modules in detail in Section 3. Then, we describe

experiments using the ICDAR 2013 table competi-

tion dataset in Section 4, showing that our method

542

Aoyagi, H., Kanazawa, T., Takasu, A., Uwano, F. and Ohta, M.

Table-structure Recognition Method Consisting of Plural Neural Network Modules.

DOI: 10.5220/0010817700003122

In Proceedings of the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2022), pages 542-549

ISBN: 978-989-758-549-4; ISSN: 2184-4313

achieved the best results compared to several existing

methods. Finally, we conclude this paper and provide

some future directions in Section 5.

2 RELATED WORK

Chi et al. proposed a graph neural network model for

recognizing the structure of tables in PDF ﬁles, named

GraphTSR (Chi et al., 2019). Their method takes a ta-

ble in PDF format as input and ﬁrst obtains cell con-

tents and its corresponding bounding box. Then, it

recognizes the table structure by predicting the rela-

tions among the obtained cells. They evaluated the

GraphTSR with the ICDAR 2013 table competition

dataset and achieved recall, precision, and F-measure

values of 0.860, 0.885, and 0.872, respectively. They

also constructed a large-scale table dataset collected

from scientiﬁc papers, named SciTSR

. This dataset

includes 12,000 tables for training and 3,000 tables

for evaluation. They demonstrated that GraphTSR

was highly effective for complicated tables. Their

method as well as our method is designed for born-

digital tables.

Paliwal et al. proposed TableNet (Paliwal et al.,

2019) that is an end-to-end deep learning model

which exploits the inherent interdependence between

the twin tasks of table detection and table structure

identiﬁcation for table images. This model consists

of two decoders: one segmentates the table region

and the other segmentates the columns in the table.

In addition, they employed a rule-based row extrac-

tion to extract the contents in cells. They evaluated

the effectiveness of their method on the ICDAR 2013

table competition dataset. As a result, their method

achieved an F-measure value of 0.915 for the table-

structure recognition and data extraction task, which

outperformed that of a deep neural network-based

method, DeepDeSRT (Schreiber et al., 2017), by 0.07

percentage points. Their method analyzes the struc-

ture of table images; however, it is not tailored to han-

dle born-digital tables in PDF documents.

Zhong et al. proposed an attention-based encoder-

dual-decoder (EDD) architecture (Zhong et al., 2020)

for table-structure recognition of table images. The

EDD consists of an encoder, a structure decoder, and

a cell decoder. The encoder captures visual features

of the image of a table. The structure decoder re-

constructs the table structure so that the cell decoder

can recognize the contents in cells. They used Pub-

TabNet

for the model training and evaluation. They

https://github.com/Academic-Hammer/SciTSR

https://github.com/ibm-aur-nlp/PubTabNet

Figure 1: Outline of the proposed method.

employed tree-edit-distance-based similarity (TEDS)

(Pawlik and Augsten, 2016) as the metric for table

recognition. Their method achieved a TEDS score

of 0.912 for simple tables and 0.854 for complex ta-

bles. These TEDS scores outperformed those of WY-

GIWYS (Deng et al., 2017) by 9.5 percentage points

and 9.9 percentage points, respectively. Their method

is also designed for table images and not for born-

digital tables.

3 TABLE-STRUCTURE

RECOGNITION

3.1 Outline of the Proposed Method

The input of the proposed method is tables in docu-

ments in PDF format. The region of the tables are lo-

cated manually before given as the input. The output

is the tables with their estimated structural informa-

tion marked up in XML format.

Figure 1 shows the outline of the proposed

method. It ﬁrst extracts tokens in a table by converting

the table in PDF to its XML ﬁle using pdfalto

as pre-

processing. It also detects ruled lines by PDFMiner

and OpenCV

. Next, it merges horizontally adjacent

tokens recursively and then estimates IRLs by the IRL

identiﬁer. Using explicit and implicit ruled lines and

token features, it merges vertically adjacent tokens re-

https://github.com/kermitt2/pdfalto

https://github.com/pdfminer/pdfminer.six

https://opencv.org

Table-structure Recognition Method Consisting of Plural Neural Network Modules

543

Figure 2: Model of the horizontal merger.

cursively and then generates cells by merging adja-

cent tokens horizontally and vertically in turns. As

postprocessing, it identiﬁes rows and columns includ-

ing multicolumn/multirow cells in the same way as

(Ohta et al., 2021).

3.2 Horizontal Merger

3.2.1 Outline

The horizontal merger continues to merge horizon-

tally adjacent tokens using their features and the sur-

rounding tokens’ features until there remain no to-

ken pairs to be merged. Its model is shown in Fig-

ure 2. The inputs of this model are 107-dimensional

two adjacent tokens’ features and 144-dimensional

surrounding tokens’ features. This model outputs

whether the adjacent tokens are to be merged. We use

the Sigmoid function as the activation of the output

layer and ReLU as the activation of the other layers.

We also use binary cross entropy as the loss function

and Adam (Kingma and Ba, 2015) as the optimizer.

We set the learning rate as 0.01 and the dropout rate

as 0.2.

3.2.2 Input Features of the Horizontal Merger

Table 1 shows two adjacent tokens’ features for the

horizontal merger, and Table 2 shows each surround-

ing token’s features.

Table 1: Two adjacent tokens’ features for the horizontal

merger.

Feature Dim.

Distance between two adjacent tokens 1

Same font or not 1

Same style or not 1

Font size (left and right tokens) 2

Numeric or not (left and right tokens) 2

Merging position 2

Table Size (height and width) 2

# of tokens in the same row and column 2

as the left token

Part of speech (POS) of left and right tokens 94

Total 107

Table 2: Each surrounding token’s features.

Feature Dim.

Position 2

Width 1

Height 1

Numeric or not 1

Total 5

In Table 1, the distance between two adjacent to-

kens is that between the right edge of the left token

and the left edge of the right token. If the token con-

sists of digits, “.,” “-,” “%,” “$,” “greater,” “smaller,”

“more,” or “less,” feature “Numeric or not” is set as 1,

and otherwise as 0. The merging position is the coor-

dinates of the middle point of the two adjacent tokens.

The table size is deﬁned by its height and width. We

obtain the part of speech (POS) of the two tokens by

the Natural Language Toolkit

Figure 3: Surrounding tokens in (a) horizontal and (b) ver-

tical merging.

https://www.nltk.org

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

544

Figure 4: Example of point clusters.

We deﬁne a token and its upper, lower, left, and

right tokens as its surrounding tokens and have two

tokens for merging, giving a total of 5 × 2 = 10 sur-

rounding tokens. Figure 3 (a) shows an example of

surrounding tokens in horizontal merging, where the

blue and green tokens are two adjacent tokens to be

inputted and the red tokens are their neighboring to-

kens. Thus, we use the ﬁve features in Table 2 for the

ten surrounding tokens, and also use POS information

of the leftmost and rightmost surrounding tokens, i.e.,

“A” and “0.87” in Figure 3 (a). We deﬁne the POS

information as a 47-dimensional one-hot vector and

thus use 5 × 10 + 47 + 47 = 144-dimensional vectors

to represent the surrounding tokens’ features.

3.3 Implicit Ruled Line Identiﬁer

3.3.1 Outline

The proposed method estimates IRLs using tokens’

features and layout. In fact, IRLs are aligned points

determined around each token, which we call point

clusters. We collect six types of point sets: the left

and right edges of a token and the middle point of

horizontally adjacent tokens for vertical IRLs, and the

upper and lower edges of a token and the middle point

of vertically adjacent tokens for horizontal IRLs. Fig-

ure 4 shows an example of such point clusters. In this

ﬁgure, green points represent the left or right points,

yellow points represent the upper or lower points, and

blue points represent the middle points between two

adjacent tokens. We then generate point clusters from

each type of point sets by the method of elastic center.

Figure 5 shows the model of the IRL identiﬁer.

The inputs of this model are 13-dimensional vectors

(cluster feature) and 100-dimensional word vectors

that are the average of the word distributed represen-

tations of the tokens related to the points in a cluster.

This word vector is acquired by word2vec (Mikolov

et al., 2013) and we use English News (2016) in

Lipzig Copora

for training the word2vec. This

model outputs whether each point cluster should be an

https://wortschatz.uni-leipzig.de/en/download

Figure 5: Model of the IRL identiﬁer.

Table 3: Features of point clusters for IRL estimation (Ohta

et al., 2021).

Feature Dim.

# of points constituting a cluster 1

# of horizontally or vertically aligned tokens 1

in a table

Existence of a ruled line 1

Direction of the alignment of points in a cluster 1

Cluster type 6

Overlap with other tokens 1

Position of a cluster 1

(relative y- or x-coordinate values)

Table size (height or width) 1

Total 13

IRL. We use the same activation functions, loss func-

tion, and optimizer as those for the horizontal merger.

We set the learning rate as 0.01 and the dropout rate

as 0.2.

3.3.2 Input Features of the Implicit Ruled Line

Identiﬁer

Table 3 shows the features of point clusters originally

determined in (Ohta et al., 2021). In Table 3, feature

“# of horizontally or vertically aligned tokens in a ta-

ble” indicates the number of horizontally (resp. verti-

cally) aligned tokens in horizontal (resp. vertical) IRL

estimation. Feature “Overlap with other tokens” is set

as 1 if the extension of an estimated IRL passes over

other tokens that are irrelevant to the estimated IRL.

Table-structure Recognition Method Consisting of Plural Neural Network Modules

545

Figure 6: Model of the vertical merger and the cell genera-

tor.

3.4 Vertical Merger

3.4.1 Outline

After the IRL estimation, the vertical merger shown

in Figure 6 continues to merge vertically adjacent

tokens using their features and the surrounding to-

kens’ features until there remain no token pairs to

be merged. Figure 3 (b) shows an example of such

two tokens and their neighboring tokens. The inputs

of this model are 119-dimensional two adjacent to-

kens’ features and 144-dimensional surrounding to-

kens’ features. This model outputs whether the adja-

cent tokens are to be merged. We use the same acti-

vation functions, loss function, and optimizer as those

for the horizontal merger. We set the learning rate as

0.01 and the dropout rate as 0.2.

3.4.2 Input Features of the Vertical Merger

We use the two adjacent tokens’ features shown in

Table 1 for vertical token merging where “left” and

“right” are substituted with “upper” and “lower,” re-

spectively. Moreover, we add the “merging direc-

tion,” “# of points constituting IRLs that exist be-

tween the two adjacent tokens,” and “textual similar-

ity between the two tokens” to those presented in Ta-

ble 1. We calculate the textual similarity by a Python

library called difﬂib. We use the same features of sur-

Figure 7: Cell estimation by the cell generator.

rounding tokens as presented in Table 2 for vertical

token merging.

3.5 Cell Generator

The cell generator continues to merge horizontally

and vertically adjacent tokens in turns using two adja-

cent tokens’ features and surrounding tokens’ features

until there remain no token pairs to be merged. The

resultant merged tokens are estimated cells. We use

the same model as that of the vertical merger shown

in Figure 6 and the same features described in Section

3.4.2 for cell generation. Note, however, that the cell

generator uses the features of left and right tokens in

horizontal merging and those of upper and lower to-

kens in vertical merging.

Figure 7 shows an example how cells are gener-

ated by the cell generator, in which red rectangles rep-

resent tokens that were not merged by the horizontal

merger nor by the vertical merger. As shown, hori-

zontally adjacent tokens such as “(PPP” and “et” are

ﬁrstly merged into one token. Then, vertically adja-

cent tokens such as “Method A” and “(PPP et” are

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

546

Table 4: Results on the ICDAR 2013 table competition dataset (G

obel et al., 2013).

Method Recall Precision F-measure

Proposed method 0.967 0.977 0.972

Ohta et al. (Ohta et al., 2021) 0.951 0.960 0.955

Nurminen (Nurminen, 2013) 0.941 0.951 0.946

1st ranked (G

obel et al., 2013)

Shigarov et al. (Shigarov et al., 2016) 0.923 0.950 0.936

GraphTSR (Chi et al., 2019) 0.860 0.885 0.872

2nd ranked (G

obel et al., 2013) 0.640 0.614 0.627

3rd ranked (G

obel et al., 2013) 0.481 0.570 0.522

Figure 8: Adjacency relation between cells.

merged into one token. Finally, two horizontally ad-

jacent tokens such as “Method A (PPP et” and “al.)”

are merged.

4 EXPERIMENTS

4.1 Dataset and Metrics

For evaluating the method performance, we used the

ICDAR 2013 table competition dataset (G

obel et al.,

2013) extracted from collected PDF documents pub-

lished by the US government and the European Union

(EU). This dataset comprised 80 tables extracted from

US documents and 76 tables extracted from EU doc-

uments, giving a total of 156 tables.

For training all the four NN models, i.e., the hori-

zontal and vertical mergers, the IRL identiﬁer, and the

cell generator, we used the same dataset as in (Ohta

et al., 2021). It consists of 209 tables, about half of

which are from the ICDAR 2013 practice dataset and

the rest are from collected research papers. Note that

we train each NN module independently.

We evaluated the table-structure recognition re-

sults based on adjacency relations between cells in

tables adopted in the ICDAR 2013 table competition

obel et al., 2012). Figure 8 shows an example of

adjacency relation between cells. In this ﬁgure, the

red circles indicate correct adjacency relations while

the blue ones indicate incorrect adjacency relations.

We can calculate recall and precision measures, as de-

ﬁned by the following equations.

Recall =

Correct adjacency relations

All adjacency relations in the ground truth

Precision =

Correct adjacency relations

All adjacency relations in the analyzed result

4.2 Experimental Results

Table 4 shows the results of table-structure recogni-

tion. As shown, the proposed method achieved the re-

call, precision, and F-measure values of 0.967, 0.977,

and 0.972, respectively, which outperforms those of

(Ohta et al., 2021) by 1.6, 1.7, and 1.7 percentage

points, respectively, and also outperforms the other

methods. Note also that the proposed method out-

performs the top-ranked participant (Nurminen) in

the ICDAR 2013 table competition by 2.6 percentage

points in F-measure.

Table 5 shows the results of the horizontal merger

at the ﬁrst iteration. As presented in the table, there

were 3,039 horizontally adjacent token pairs to be

merged and 85 pairs not to be merged. The pairs

to be merged were much more than those not to be

merged because we set a relatively small threshold for

the distance between them when selecting the initial

token pairs. As seen here, the horizontal merger could

merge 3,031 out of 3,039 token pairs that should be

merged while it mistakenly merged 19 out of 85 to-

ken pairs that should not be merged.

Figure 9 shows a part of an incorrectly analyzed

table by the horizontal merger where “Less than” and

Figure 9: Failed example of horizontal token merging in the

dataset (G

obel et al., 2013).

Table-structure Recognition Method Consisting of Plural Neural Network Modules

547

Figure 10: Successfully analyzed table in the dataset (G

obel et al., 2013).

Table 5: Results of merging horizontally adjacent tokens.

Predicted

Merged Not merged

Ground truth Merged 3,031 8

Not merged 19 66

“$10,000-” were mistakenly merged into one token

because the distance between them is very small and

“Less than $10,000-” appears to make sense. Note

that “Less than” is a token generated by correctly

merging “Less” and “than.”

Table 6 shows the results of IRL estimation. As

shown, there were 13,132 IRL point clusters and

5,000 non-IRL point clusters. Because the non-

IRL samples were fewer than the IRL samples, the

non-IRL samples were augmented by SMOTEENN

(Batista et al., 2004), when training the IRL identi-

ﬁer. The errors where the non-IRL point clusters were

mistakenly estimated as IRLs were more than the er-

rors where the IRL point clusters were mistakenly es-

timated as non-IRLs.

Table 7 shows the results of the cell generator at

the ﬁrst iteration. In contrast to Table 5, there were

much more token pairs not to be merged (25,539) than

those to be merged (1,656) in Table 7 partly because

these tokens were remained tokens after both horizon-

tal and vertical token merging. Note also that these

token pairs consist of both horizontally and vertically

adjacent token pairs found in the identical token set

generated by the vertical merger. As seen in Table 7,

less than one-ﬁfteenth of the pairs should be merged

and the rest should not, being also imbalanced. How-

ever, we did not adjust the imbalance of samples be-

cause the errors where the token pairs to be merged

are mistakenly not merged are not critical compared

with the errors where the token pairs not to be merged

are mistakenly merged. As presented in the table, the

number of mistakenly merged token pairs was only

55, i.e., 0.22% of 25,539 token pairs not to be merged

while the number of token pairs failing to be merged

was 376, i.e., 23% of 1,656 token pairs to be merged.

Figure 10 shows a part of a correctly analyzed ta-

ble by the proposed method where the table has plural

large cells ﬁlled with a long passage. In analysis of

this kind of tables, the horizontal merger is especially

Table 6: Results of IRL estimation.

Predicted

IRL Non-IRL

Ground truth IRL 12,564 568

Non-IRL 880 4,120

Table 7: Results of merging tokens by the cell generator.

Predicted

Merged Not merged

Ground truth Merged 1,280 376

Not merged 55 25,484

effective in correct IRL estimation because each word

is usually recognized as a token, thereby generating

many non-IRL point clusters around it unless these

horizontally adjacent tokens are merged. The hori-

zontal merger successfully merged horizontally adja-

cent tokens in the table repeatedly before IRL estima-

tion, which led to correct cell estimation by the cell

generator.

5 CONCLUSION

We proposed a table-structure recognition method us-

ing plural NN modules in this paper. The proposed

method consists of four NN modules: one merges to-

kens in a table horizontally and one vertically, one es-

timates IRLs that are necessary to separate cells but

undrawn, and the last ﬁnally estimates cells by merg-

ing the tokens. We analyzed 156 tables in the ICDAR

2013 table-structure competition dataset, for which

the resultant recall, precision, and F-measure values

were 0.967, 0.977, and 0.972, respectively. These

results outperformed those of our previous method

(Ohta et al., 2021) in every metric and by 1.7 per-

centage points in F-measure, vastly outperforming the

other methods, including the top-ranked participant

Nurminen (Nurminen, 2013) in that competition.

In future work, we plan to apply our table-

structure recognition method to other datasets such as

SciTSR (Chi et al., 2019) and to table image anal-

ysis problems tackled at ICDAR 2019 (Gao et al.,

2019). We also plan to introduce a table region detec-

tion module into our method by using state-of-the-art

ICPRAM 2022 - 11th International Conference on Pattern Recognition Applications and Methods

548

table detection algorithms such as (Riba et al., 2019)

and (Prasad et al., 2020). Moreover, we want to de-

velop an automatic graph generation application us-

ing the tables analyzed by our method.

ACKNOWLEDGEMENTS

This work was supported by a JSPS Grant-in-Aid

for Scientiﬁc Research (C) (18K11989), the Cross-

ministerial Strategic Innovation Promotion Program

(SIP) Second Phase, “Big-data and AI-enabled Cy-

berspace Technologies” by NEDO, and ROIS NII

Open Collaborative Research 2021 (21FC04).

REFERENCES

Batista, G., Prati, R., and Monard, M. (2004). A study

of the behavior of several methods for balancing ma-

chine learning training data. SIGKDD Explor. Newsl.,

6(1):20–29.

Chi, Z., Huang, H., Xu, H., Yu, H., Yin, W., and Mao,

X. (2019). Complicated table structure recognition.

CoRR, abs/1908.04729.

Deng, Y., Kanervisto, A., Ling, J., and Rush, M. A. (2017).

Image-to-markup generation with coarse-to-ﬁne at-

tention. In 34th International Conference on Machine

Learning (ICML), volume 70, pages 980–989.

Gao, L., Huang, Y., D

ejean, H., Meunier, J.-L., Yan, Q.,

Fang, Y., Kleber, F., and Lang, E. (2019). ICDAR

2019 competition on table detection and recognition

(cTDaR). In 2019 15th International Conference on

Document Analysis and Recognition (ICDAR), pages

1510–1515.

obel, M., Hassan, T., Oro, E., and Orsi, G. (2013). IC-

DAR 2013 table competition. In 2013 12th Interna-

tional Conference on Document Analysis and Recog-

nition (ICDAR), pages 1449–1453.

obel, M., Hassan, T., Oro, E., and Orsi, G. (2012). A

methodology for evaluating algorithms for table un-

derstanding in pdf documents. In ACM Symposium

on Document Engineering 2012 (DocEng ’12), pages

45–48.

Kingma, P. D. and Ba, J. (2015). Adam: A method for

stochastic optimization. In 3rd International Confer-

ence on Learning Representations (ICLR ’15).

Mikolov, T., Chen, K., Corrado, G., and Dean, J. (2013).

Efﬁcient estimation of word representations in vector

space. In the Workshop at 1st International Confer-

ence on Learning Representations.

Nurminen, A. (2013). Algorithmic extraction of data in ta-

bles in pdf documents. Master’s thesis, Tampere Uni-

versity of Technology.

Ohta, M., Yamada, R., Kanazawa, T., and Takasu, A.

(2019). A cell-detection-based table-structure recog-

nition method. In ACM Symposium on Document En-

gineering 2019 (DocEng ’19), Article 27, 4 pages.

Ohta, M., Yamada, R., Kanazawa, T., and Takasu, A.

(2021). Table-structure recognition method using neu-

ral networks for implicit ruled line estimation and cell

estimation. In ACM Symposium on Document Engi-

neering 2021 (DocEng ’21), Article 23, 7 pages.

Paliwal, S., Vishwanath, D., Rahul, R., Sharma, M., and

Vig, L. (2019). Tablenet: Deep learning model for

end-to-end table detection and tabular data extraction

from scanned document images. In 2019 15th Interna-

tional Conference on Document Analysis and Recog-

nition (ICDAR), pages 128–133.

Pawlik, M. and Augsten, N. (2016). Tree edit distance:

Robust and memory-efﬁcient. Information Systems,

56:157–173.

Prasad, D., Gadpal, A., Kapadni, K., Visave, M., and Sul-

tanpure, K. (2020). CascadeTabNet: An approach for

end to end table detection and structure recognition

from image-based documents. In 2020 IEEE/CVF

Conference on Computer Vision and Pattern Recog-

nition Workshops (CVPRW), pages 2439–2447.

Riba, P., Dutta, A., Goldmann, L., Forn

es, A., Ramos,

O., and Llad

os, J. (2019). Table detection in invoice

documents by graph neural networks. In 2019 15th

International Conference on Document Analysis and

Recognition (ICDAR), pages 122–127.

Schreiber, S., Agne, S., Wolf, I., Dengel, A., and Ahmed,

S. (2017). Deepdesrt: Deep learning for detection and

structure recognition of tables in document images. In

2017 14th IAPR International Conference on Docu-

ment Analysis and Recognition (ICDAR), volume 1,

pages 1162–1167.

Shigarov, A., Mikhailov, A., and Altaev, A. (2016). Con-

ﬁgurable table structure recognition in untagged pdf

documents. In ACM Symposium on Document Engi-

neering 2016 (DocEng ’16), pages 119–122.

Zhong, X., ShaﬁeiBavani, E., and Yepes, J. A. (2020).

Image-based table recognition: Data, model, and eval-

uation. In 16th European Conference on Computer

Vision (ECCV ’20), pages 564–580.

Table-structure Recognition Method Consisting of Plural Neural Network Modules

549