Deep Learning for Effective Classiﬁcation and Information Extraction of

Financial Documents

Valentin-Adrian Serbanescu

and Maruf Dhali

Department of Artiﬁcial Intelligence, Bernoulli Institute, University of Groningen, 9747 AG Groningen, The Netherlands

Keywords:

Deep Learning, Classiﬁcation, Information Extraction, Financial Documents, Optical Character Recognition,

CNN, RoBERTa, LayoutLMv3, GraphDoc, Document Processing.

Abstract:

The ﬁnancial and accounting sectors are encountering increased demands to effectively manage large vol-

umes of documents in today’s digital environment. Meeting this demand is crucial for accurate archiving,

maintaining efﬁciency and competitiveness, and ensuring operational excellence in the industry. This study

proposes and analyzes machine learning-based pipelines to effectively classify and extract information from

scanned and photographed ﬁnancial documents, such as invoices, receipts, bank statements, etc. It also ad-

dresses the challenges associated with ﬁnancial document processing using deep learning techniques. This

research explores several models, including LeNet5, VGG19, and MobileNetV2 for document classiﬁcation

and RoBERTa, LayoutLMv3, and GraphDoc for information extraction. The models are trained and tested on

ﬁnancial documents from previously available benchmark datasets and a new dataset with ﬁnancial documents

in Romanian. Results show MobileNetV2 excels in classiﬁcation tasks (with accuracies of 99.24% with data

augmentation and 93.33% without augmentation), while RoBERTa and LayoutLMv3 lead in extraction tasks

(with F1-scores of 0.7761 and 0.7426, respectively). Despite the challenges posed by the imbalanced dataset

and cross-language documents, the proposed pipeline shows potential for automating the processing of ﬁnan-

cial documents in the relevant sectors.

1 INTRODUCTION

In today’s digital era, the rapid growth of data high-

lights the need for efﬁcient processing of ﬁnancial

documents — especially in ﬁnance and accounting,

where manually handling invoices and receipts can

be labor-intensive and prone to errors. This ar-

ticle aims to develop a machine-learning pipeline

that classiﬁes and extracts critical information from

scanned ﬁnancial documents. It will utilize advanced

techniques such as Convolutional Neural Networks

(CNNs), Graph Attention Networks (GATs), Trans-

formers, and Optical Character Recognition (OCR).

The proposed system addresses the complexities

of various document formats, enhancing the accu-

racy and speed of ﬁnancial data processing. With

the growing demand for automation in large organi-

zations like banks and accounting ﬁrms, this article

seeks to optimize ﬁnancial operations.

This work presents a machine learning pipeline

capable of accurately classifying and extracting valu-

https://orcid.org/0009-0003-6678-6536

https://orcid.org/0000-0002-7548-3858

able information from diverse ﬁnancial documents,

laying the groundwork for future industrial appli-

cations. All materials and scripts are available on

GitHub

1.1 Related Works

Document classiﬁcation in the ﬁnancial sector has be-

come crucial for processing large volumes of docu-

ments like invoices and bank statements. Initially re-

liant solely on Optical Character Recognition (OCR),

the ﬁeld has evolved with machine learning advance-

ments, mainly through Convolutional Neural Net-

works (CNNs). In the study by (Chen and Blostein,

2007), the authors emphasized the limitations of tradi-

tional classiﬁcation methods relying solely on OCR.

Later, (Kang et al., 2014) demonstrated the effective-

ness of CNNs in handling the structural hierarchies

and spatial relationships of document images, show-

ing a signiﬁcant improvement over earlier techniques.

(Rusi

nol et al., 2014) introduced a multimodal ap-

GitHub Repository (private - available upon request):

DL-ﬁnancial-doc-classiﬁcation-extraction

Serbanescu, V.-A. and Dhali, M.

Deep Learning for Effective Classiﬁcation and Information Extraction of Financial Documents.

DOI: 10.5220/0013261000003905

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 749-756

ISBN: 978-989-758-730-6; ISSN: 2184-4313

749

proach combining CNNs for visual analysis and OCR

for text extraction, advancing multi-page document

classiﬁcation. (Harley et al., 2015) further explored

deep CNNs for document classiﬁcation, underscoring

their capability to manage high visual variability in

document categories. More recent research by (Dong

and Li, 2020) applied a simpliﬁed CNN architecture

speciﬁcally to ﬁnancial documents, like invoices and

bank receipts, demonstrating practical applications in

the ﬁnance sector. On the other hand, (Lehtonen

et al., 2020) explored the k-Nearest Neighbors (kNN)

algorithm for document classiﬁcation, pointing out

the potential of simpler models, while (

Omer Ar-

slan and Uymaz, 2022) provided insights into CNN-

based architectures like LeNet-5, VGG-19, and Mo-

bileNetV2, highlighting techniques like padding and

data augmentation for enhancing classiﬁcation accu-

racy.

Considerable progress has been achieved in doc-

ument information extraction with the advancement

of pre-trained models. (Liu et al., 2019) introduced

RoBERTa, an enhanced version of BERT, which

showed superior performance on various benchmarks

by optimizing training methods and data. (Majumder

et al., 2020) proposed a representation learning ap-

proach for form-like documents, signiﬁcantly im-

proving extraction performance across multiple do-

mains by generating extraction candidates based on

target ﬁelds. Another critical study by (Oral et al.,

2020) tackled the challenge of extracting informa-

tion from visually complex banking documents using

deep learning algorithms and neural word represen-

tations such as FastText, ELMo, and BERT. The au-

thors noted a signiﬁcant 10% improvement in named

entity recognition and relation extraction. (Huang

et al., 2022) introduced LayoutLMv3, a multimodal

pre-trained model for Document AI, combining text

and image masking to align modalities, leading to

state-of-the-art performance across various document

understanding tasks without the need for CNN-based

image feature extraction. (Ha and Hor

ak, 2022) de-

veloped the OCRMiner system, designed speciﬁcally

for extracting information from scanned documents

like invoices, showing success rates of over 88% for

Czech and English invoices. Finally, (

Simsa

et al., 2023) presented the DocILE benchmark, using

RoBERTa and LayoutLMv3 for essential information

extraction tasks, further contributing to advancements

in document AI.

Chargrids: The introduction of chargrids by (Katti

et al., 2018) added a new dimension to document un-

derstanding, converting documents into a 2D grid of

characters for processing with CNNs. This approach

preserves the spatial layout of the document and im-

proves semantic segmentation and information ex-

traction, particularly from layout-rich documents like

invoices (see Figure 1). By representing documents

in this format, the chargrid model captures both tex-

tual and spatial information, which proves beneﬁcial

for information extraction, particularly in layout-rich

documents like invoices. In their experiments, the au-

thors demonstrated that chargrid signiﬁcantly outper-

forms traditional text-only or image-based methods

in extracting structured information from documents

with complex layouts. Based on these ﬁndings, we in-

corporate chargrid into our model training process for

information extraction, aiming to leverage its ability

to improve spatial and layout understanding in docu-

ment processing tasks.

Figure 1: Example of chargrid created on the basis of a ﬁ-

nancial document. Figure adapted from the study of (Katti

et al., 2018).

Data Augmentation: Data augmentation tech-

niques play a vital role in improving model perfor-

mance. (Perez and Wang, 2017) explored various

augmentation methods such as cropping, rotating, and

GAN-based style transfer. Similarly, (Mikołajczyk-

Bareła and Grochowski, 2018) proposed blending

content and style through image style transfer, which

proved particularly effective in tasks like medical

imaging. (Shorten and Khoshgoftaar, 2019) cat-

egorized augmentation methods into data warping

and oversampling techniques, such as SMOTE and

GANs, while (Yang et al., 2023) provided a com-

prehensive taxonomy of augmentation methods, high-

lighting the beneﬁts of AutoAugment for optimizing

policies. These techniques, particularly traditional

ones, are essential for improving the performance of

models trained on ﬁnancial documents, as demon-

strated by previous research. We will only use tra-

ditional augmentation techniques to simulate realis-

tic variations of ﬁnancial documents based on con-

clusions from related works.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

750

2 METHODS

2.1 Classiﬁcation

In the classiﬁcation task, this study uses two datasets:

a new proposed dataset with Romanian Financial

Documents (RFD) comprising seven classes (invoice,

receipt, accompanying notice of wares, bank state-

ment, return receipt, payment disposition, and collec-

tion disposition) and the RVL-CDIP Dataset (Harley

et al., 2015). The RFD dataset contains 896 samples

distributed across these classes (for example, see Fig-

ure 6 in the Appendix), revealing notable class im-

balances, where certain document classes are under-

represented compared to others, as illustrated in Fig-

ure 2. To address these discrepancies, traditional data

augmentation techniques, including ﬂipping, rotating,

translation, scaling, random cropping, brightness ad-

justment, and Gaussian noise, are applied to increase

diversity within each class. These augmentation tech-

niques lead to a more balanced distribution of sam-

ples, enhancing the dataset’s diversity, as seen in Fig-

ure 3, and ultimately improving the model’s general-

ization ability.

The RVL-CDIP Dataset, a subset of the IIT-CDIP

Test Collection, contains 400,000 grayscale images in

16 document categories (for example, see Figure 7 in

the Appendix). From these categories, seven similar

ﬁnancial and business document types are selected to

align with the CD, providing a consistent foundation

for evaluating classiﬁcation performance across var-

ious document types. This alignment allows a more

comprehensive comparison of model effectiveness on

different document distributions and class structures

across the two datasets.

The models used for classiﬁcation include various

CNN architectures:

• CNN: Three convolutional layers followed by

fully connected layers, based on the works of

(Dong and Li, 2020), (Harley et al., 2015), and

(Kang et al., 2014).

• LeNet5: A classic two-layer convolutional archi-

tecture from (Lecun et al., 1998).

• VGG19: A modiﬁed version of the VGG19 net-

work, as outlined by (Simonyan and Zisserman,

2015), with single-channel input.

• MobileNetV2: Based on (Sandler et al., 2019),

using depthwise separable convolutions for re-

duced complexity.

Experiments. Four experiments are conducted:

two using the RFD dataset and two with a subset of

Figure 2: Number of samples per ﬁnancial document class

in RVL-CDIP dataset.

Figure 3: Number of samples per ﬁnancial document class

after data augmentation.

RVL-CDIP. Each experiment employed 5-fold cross-

validation, with an 80/20 split for training+validation

and testing and another 80/20 split for training and

validation. Results were generated from validation

and testing, with testing accuracy calculated using

models trained on combined training and valida-

tion sets. Evaluation metrics included mean accu-

racy, standard deviation, and testing accuracy across

folds. The experiments aimed to assess the impact

of data augmentation on the RFD dataset, evaluate

pre-trained models on RVL-CDIP when applied to the

augmented RFD dataset, measure model performance

on the RVL-CDIP subset, and test the performance

of models pre-trained on the augmented RFD dataset

when applied to RVL-CDIP.

Deep Learning for Effective Classiﬁcation and Information Extraction of Financial Documents

751

Figure 4: Test accuracy of various models on different datasets for the classiﬁcation task (see also Tables 1, 2, 3, and 4).

Table 1: Classiﬁcation accuracy using the RFD dataset with and without data augmentation.

Models Test (Aug) Val Mean (Aug) Val Std (Aug) Test (No Aug) Val Mean (No Aug) Val Std (No Aug)

CNN 95.89 95.16 0.98 92.78 92.87 1.01

LeNet5 92.94 89.25 2.43 85.56 81.28 4.37

MobileNetV2 (pretrained) 99.24 93.10 8.78 93.33 94.41 4.12

MobileNetV2 95.28 95.38 1.59 65.56 86.17 7.80

VGG19 (pretrained) 87.98 73.17 31.52 91.11 89.52 2.39

VGG19 16.44 14.51 1.40 47.22 47.33 4.56

2.2 Information Extraction

The information extraction task focuses on Key Infor-

mation Localization and Extraction (KILE) and Line

Item Recognition (LIR). Two datasets are used: the

RFD dataset of manually annotated ﬁnancial docu-

ments (35 documents across seven classes) and the

DocILE Benchmark Dataset (

Simsa et al.,

2023). The DocILE dataset comprises 6.7k annotated

business documents, 100k synthetic documents, and

1 million unlabeled documents. The RFD dataset in-

cludes ﬁelds like invoice numbers, item descriptions,

and total values. These datasets test the ability to ex-

tract essential information from ﬁnancial documents.

The models used for information extraction in-

clude:

• RoBERTa: A transformer model for text-based

tasks, as described by (Liu et al., 2019).

• LayoutLMv3: A multimodal model handling both

text and layout, based on the work of (Huang

et al., 2022).

• GraphDoc: A graph-based model that integrates

text, layout, and visual features, as described by

(Wang et al., 2023).

Experiments: Two experiments were conducted:

one on the DocILE dataset and the other on the

RFD dataset. The ﬁrst experiment used 5-fold cross-

validation on DocILE, with models evaluated on F1-

score, average precision (AP), precision, and recall.

This experiment aimed to compare results with those

of (

Simsa et al., 2023) and to assess model ro-

bustness by combining KILE and LIR tasks for over-

all performance, unlike the separate evaluations in the

original study.

In the second experiment, models trained on En-

glish documents were tested on a manually annotated

RFD dataset, translated from Romanian to English.

This experiment used F1-score, AP, precision, and re-

call metrics for both KILE and LIR tasks to evaluate

model performance on novel, non-English data.

3 RESULTS

This section presents the classiﬁcation results of the

RFD and RVL-CDIP datasets (see Figure 4 for the

plot). For the information extraction tasks, focusing

on the DocILE benchmark and the RFD dataset, see

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

752

Figure 5 in the Appendix for the results plot. Detailed

results are discussed in the following subsections.

3.1 Classiﬁcation Task

RFD Dataset: The results for the classiﬁcation task

on the RFD dataset with and without data augmenta-

tion are shown in Table 1. The CNN model achieved

the highest accuracy on unseen data (95.89%)

with augmentation, while MobileNetV2 (pretrained)

achieved the highest accuracy without augmentation

(93.33%). Data augmentation signiﬁcantly improved

model performance across most models. Pretrained

models, particularly MobileNetV2 and VGG19, con-

sistently outperformed their non-pretrained counter-

parts.

RVL-CDIP Dataset: The results for the RVL-

CDIP dataset are presented in Table 2. MobileNetV2

(pretrained) achieved the highest accuracy (76.34%),

while CNN and LeNet5 followed with lower perfor-

mance. Pretrained models performed signiﬁcantly

better than non-pretrained ones, highlighting the ben-

eﬁt of transfer learning.

RFD Dataset with RVL-CDIP Pretraining: The

results for the RFD dataset with models pretrained on

the RVL-CDIP dataset are in Table 3. MobileNetV2

(pretrained) and VGG19 (pretrained) achieved high

accuracy, showing that models pretrained on RVL-

CDIP can generalize well to the RFD dataset.

RVL-CDIP Dataset with RFD Pretraining: Ta-

ble 4 shows that MobileNetV2 (pretrained) achieved

the highest accuracy at 74.65%, followed by VGG19

(pretrained) at 62.98%. The CNN performed well

with 64.21% accuracy. LeNet5 and non-pretrained

models underperformed, with VGG19 scoring the

lowest at 14.29%. Pretrained models generalize better

to RVL-CDIP.

3.2 Information Extraction Task

DocILE Dataset: The results for the combined

KILE and LIR tasks on the DocILE Benchmark

Dataset are presented in Table 5. RoBERTa achieved

the highest F1-score (0.7761), while LayoutLMv3

performed slightly lower. GraphDoc showed weaker

performance compared to the other models.

RFD Dataset: The results for the RFD dataset

for KILE and LIR tasks are presented in Table 6.

RoBERTa showed strong performance in the KILE

task but weaker in the LIR task. LayoutLMv3 per-

formed well across both functions, while GraphDoc

underperformed in comparison.

4 CONCLUSIONS

Modern architectures like MobileNetV2 (pretrained)

and CNN, combined with data augmentation, de-

livered the best results for the classiﬁcation tasks.

Pretrained models consistently outperformed non-

pretrained ones, emphasizing the importance of trans-

fer learning. In document information extraction

tasks, RoBERTa and LayoutLMv3 proved the most

effective, with RoBERTa excelling in token-level pre-

dictions. LayoutLMv3 performed well, particularly

for structured data extraction like line items. Graph-

Doc struggled in both tasks, indicating limitations

for complex document processing. MobileNetV2

and CNN are best suited for classiﬁcation, while

RoBERTa and LayoutLMv3 excel in information ex-

traction.

Hardware limitations impacted this work, particu-

larly during model training for complex models like

RoBERTa and LayoutLMv3. The RFD dataset used

for classiﬁcation was relatively small, which may

have affected the models’ ability to generalize across

various document types. Furthermore, the annota-

tions for information extraction were inadequate, and

translation errors in the RFD dataset could have fur-

ther hindered model performance. By addressing

these challenges through the use of larger datasets,

improved annotations, and enhanced computational

resources, we are likely to achieve more accurate

models.

Future research should concentrate on developing

a benchmark speciﬁcally for classifying ﬁnancial doc-

uments, as existing datasets may not adequately re-

ﬂect the unique features of these documents. Ad-

ditionally, creating multilingual models is crucial,

particularly for processing languages like Romanian,

to enhance the effectiveness of document extraction

systems in various linguistic contexts. Expanding

datasets in both size and diversity will be essential

for improving the generalization and performance of

models in classiﬁcation and information extraction

tasks.

In conclusion, this research proposes and ana-

lyzes machine learning pipelines for classifying and

extracting information from ﬁnancial documents. It

demonstrates promising results in automating docu-

ment processing and paves the way for further devel-

opment in industrial applications.

Deep Learning for Effective Classiﬁcation and Information Extraction of Financial Documents

753

Table 2: Classiﬁcation accuracy using the RVL-CDIP dataset.

Models Test Accuracy Val Mean Val Std

CNN 65.59 63.03 0.60

LeNet5 61.75 59.58 0.96

MobileNetV2 (pretrained) 76.34 72.64 2.49

MobileNetV2 64.36 58.94 1.18

VGG19 (pretrained) 53.15 51.89 20.14

VGG19 14.29 13.12 1.06

Table 3: Classiﬁcation accuracy using the RFD dataset with models pretrained on the RVL-CDIP dataset.

Models Test Accuracy Val Mean Val Std

CNN 94.06 93.63 0.96

LeNet5 94.22 91.15 1.20

MobileNetV2 (pretrained) 91.93 97.21 1.48

MobileNetV2 95.88 96.76 1.17

VGG19 (pretrained) 98.17 94.62 2.87

VGG19 16.44 16.46 1.91

Table 4: Classiﬁcation accuracy using the RVL-CDIP dataset with models pretrained on the RFD dataset.

Models Test Accuracy Val Mean Val Std

CNN 64.21 62.08 1.21

LeNet5 47.16 48.13 3.22

MobileNetV2 (pretrained) 74.65 73.59 1.45

MobileNetV2 53.76 55.37 0.73

VGG19 (pretrained) 62.98 60.40 1.13

VGG19 14.29 11.96 0.37

Table 5: Information extraction results (F1-score, average precision, precision, recall) for the 5-fold cross-validation on the

DocILE dataset.

Models F1-score average precision precision recall

RoBERTa 0.7761 ± 0.0268 0.8194 ± 0.0204 0.8180 ± 0.0252 0.7650 ± 0.0253

LayoutLMv3 0.7426 ± 0.0198 0.7878 ± 0.0198 0.7752 ± 0.0260 0.7324 ± 0.0148

GraphDoc 0.4497 ± 0.0043 0.4974 ± 0.0059 0.5133 ± 0.0051 0.4402 ± 0.0060

Table 6: Information extraction results for KILE and LIR tasks on the RFD dataset, including F1-score, precision, recall, and

average precision.

Models F1-score (KILE) Avg. precision (KILE) precision (KILE) recall (KILE) F1-score (LIR) Avg. precision (LIR) precision (LIR) recall (LIR)

RoBERTa 0.5101 0.0133 0.5101 0.5101 0.2439 0.1522 0.2439 0.2439

LayoutLMv3 0.4973 0.0152 0.4973 0.4973 0.3396 0.2234 0.3396 0.3396

GraphDoc 0.3033 0.0200 0.3033 0.3033 0.2052 0.1532 0.2052 0.2052

ACKNOWLEDGMENTS

We thank Vamadi Contab, the accounting company in

Romania (https://vamadi.ro), for providing the ﬁnan-

cial documents that enabled the practical validation of

our models. Finally, we appreciate the H

abr

ok HPC

Cluster for providing computational resources, with-

out which the article would not have reached its cur-

rent depth and scope.

REFERENCES

Chen, N. and Blostein, D. (2007). A survey of document

image classiﬁcation: problem statement, classiﬁer ar-

chitecture and performance evaluation. In Interna-

tional Journal of Document Analysis and Recognition

(IJDAR), volume 10, pages 1–16.

Dong, J. and Li, X. (2020). An image classiﬁcation algo-

rithm of ﬁnancial instruments based on convolutional

neural network. In Traitement du Signal, volume 37,

pages 1055–1060.

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

754

Ha, H. and Hor

ak, A. (2022). Information extraction from

scanned invoice images using text analysis and layout

features. In Signal Processing: Image Communica-

tion, volume 102, page 116601.

Harley, A. W., Ufkes, A., and Derpanis, K. G. (2015). Eval-

uation of deep convolutional nets for document image

classiﬁcation and retrieval. In 2015 13th International

Conference on Document Analysis and Recognition

(ICDAR), pages 991–995.

Huang, Y., Lv, T., Cui, L., Lu, Y., and Wei, F. (2022). Lay-

outlmv3: Pre-training for document ai with uniﬁed

text and image masking.

Kang, L., Kumar, J., Ye, P., Li, Y., and Doermann, D. S.

(2014). Convolutional neural networks for document

image classiﬁcation. In 2014 22nd International Con-

ference on Pattern Recognition, pages 3168–3172.

Katti, A. R., Reisswig, C., Guder, C., Brarda, S., Bickel,

S., H

ohne, J., and Faddoul, J. B. (2018). Chargrid:

Towards understanding 2d documents. arXiv preprint.

Lecun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998).

Gradient-based learning applied to document recogni-

tion. In Proceedings of the IEEE, volume 86, pages

2278–2324.

Lehtonen, R., Nevalainen, P., and Murtoj

arvi, M. (2020).

Automated classiﬁcation of receipts and invoices

along with document extraction. In University of

Turku.

Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D.,

Levy, O., Lewis, M., Zettlemoyer, L., and Stoyanov,

V. (2019). Roberta: A robustly optimized bert pre-

training approach.

Majumder, B. P., Potti, N., Tata, S., Wendt, J. B., Zhao,

Q., and Najork, M. (2020). Representation learning

for information extraction from form-like documents.

In Proceedings of the 58th Annual Meeting of the As-

sociation for Computational Linguistics, pages 6495–

6504.

Mikołajczyk-Bareła, A. and Grochowski, M. (2018). Data

augmentation for improving deep learning in image

classiﬁcation problem. In IIPHDW, pages 117–122.

Oral, B., Emekligil, E., Arslan, S., and Eryi

git, G. (2020).

Information extraction from text intensive and visu-

ally rich banking documents. In Information Process-

ing and Management, volume 57, page 102361.

Perez, L. and Wang, J. (2017). The effectiveness of data

augmentation in image classiﬁcation using deep learn-

ing.

Rusi

nol, M., Frinken, V., Karatzas, D., Bagdanov, A. D.,

and Llad

os, J. (2014). Multimodal page classiﬁcation

in administrative document image streams. In Interna-

tional Journal on Document Analysis and Recognition

(IJDAR), volume 17, pages 331–341.

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and

Chen, L.-C. (2019). Mobilenetv2: Inverted residuals

and linear bottlenecks.

Shorten, C. and Khoshgoftaar, T. (2019). A survey on image

data augmentation for deep learning. In Journal of Big

Data.

Simonyan, K. and Zisserman, A. (2015). Very deep convo-

lutional networks for large-scale image recognition.

Wang, Y., Du, J., Ma, J., Hu, P., Zhang, Z., and Zhang,

J. (2023). Ustc-iﬂytek at docile: A multi-modal ap-

proach using domain-speciﬁc graphdoc. In Confer-

ence and Labs of the Evaluation Forum.

Yang, S., Xiao, W., Zhang, M., Guo, S., Zhao, J., and Shen,

F. (2023). Image data augmentation for deep learning:

A survey.

Omer Arslan and Uymaz, S. A. (2022). Classiﬁcation of in-

voice images by using convolutional neural networks.

In Journal of Advanced Research in Natural and Ap-

plied Sciences, volume 8, pages 8–25. C¸ anakkale On-

sekiz Mart University.

Simsa,

Sulc, M., U

r, M., Patel, Y., Hamdi, A.,

Koci

an, M., Matas, J., Doucet, A., Coustaty, M., and

Karatzas, D. (2023). Docile benchmark for document

information localization and extraction.

APPENDIX

Figure 5: F1-scores of different models tested on different

datasets for the information extraction task.

Deep Learning for Effective Classiﬁcation and Information Extraction of Financial Documents

755

Figure 6: Example documents and classes from RFD dataset used for the image classiﬁcation task.

Figure 7: Example documents and classes from the RVL-CDIP Dataset, ﬁgures taken from the study of (Harley et al., 2015).

ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods

756