EHDI: Enhancement of Historical Document Images via Generative

Adversarial Network

Abir Fathallah

1,2 a

, Mounim A. El-Yacoubi

and Najoua Essoukri Ben Amara

Universit

e de Sousse, Institut Sup

erieur de l’Informatique et des Techniques de Communication,

LATIS - Laboratory of Advanced Technology and Intelligent Systems, 4023, Sousse, Tunisia

Samovar, CNRS, T

ecom SudParis, Institut Polytechnique de Paris, 9 rue Charles Fourier, 91011 Evry Cedex, France

Universit

e de Sousse, Ecole Nationale d’Ing

enieurs de Sousse,

LATIS-Laboratory of Advanced Technology and Intelligent Systems, 4023, Sousse, Tunisia

Keywords:

Historical Documents, Document Enhancement, Degraded Documents, Generative Adversarial Networks.

Abstract:

Images of historical documents are sensitive to the signiﬁcant degradation over time. Due to this degrada-

tion, exploiting information contained in these documents has become a challenging task. Consequently, it is

important to develop an efﬁcient tool for the quality enhancement of such documents. To address this issue,

we present in this paper a new modelknown as EHDI (Enhancement of Historical Document Images) which

is based on generative adversarial networks. The task is considered as an image-to-image conversion process

where our GAN model involves establishing a clean version of a degraded historical document. EHDI implies

a global loss function that associates content, adversarial, perceptual and total variation losses to recover global

image information and generate realistic local textures. Both quantitative and qualitative experiments demon-

strate that our proposed EHDI outperforms signiﬁcantly the state-of-the-art methods applied to the widespread

DIBCO 2013, DIBCO 2017, and H-DIBCO 2018 datasets. Our suggested model is adaptable to other docu-

ment enhancement problems, following the results across a wide range of degradations. Our code is available

at https://github.com/Abir1803/EHDI.git.

1 INTRODUCTION

Historical Arabic Documents (HADs) are a valuable

part of cultural heritage, but access to them is of-

ten limited due to inadequate storage conditions. To

be understood automatically by machine vision, dig-

ital historical documents must be transcribed into a

readable form, as they are not readily processed in

their original form. These documents often suffer

from various types of degradation. The restoration

of historical documents can be complicated by the

presence of watermarks, stamps, or annotations, es-

pecially when these forms of degradation occur in

the text itself. This is particularly challenging when

the stain color is similar to or more intense than the

font color of the document. Document processing,

which involves transcribing digital historical docu-

ments into a readable form, can be done using a com-

puter vision tool or by a human being. In recent

years, the development of various public databases

has led to a signiﬁcant expansion of document pro-

cessing. The processing of historical documents is

https://orcid.org/0000-0003-0433-1029

a very challenging task and it is not always efﬁcient

due to the poor quality of manuscripts. These doc-

uments can be affected by various types of damage,

such as wrinkles, dust damage, nutrition stains, and

discolored sunspots, which can hinder their process-

ing efﬁciency (Zamora-Mart

ınez et al., 2007). Degra-

dation may also occur in the scanned documents due

to poor scanning conditions related to the use of the

smartphone camera (shadow (Finlayson et al., 2002),

blurring (Chen et al., 2011), varying light conditions,

warping, etc.). In addition, several documents are

sometimes infested with stamps, watermarks, or an-

notations. In this paper, a new document enhance-

ment model is evolved for enhancing degraded docu-

ments to provide a cleaned-up version. Speciﬁcally,

we consider the document enhancement task as a

GAN-based image-to-image converter process. This

paper proposes a new GAN architecture speciﬁcally

designed to improve the clarity of images of histori-

cal documents. In contrast to previous methods, this

approach aims to simultaneously remove noise and

watermarks while preserving the quality of the text.

The ultimate goal is to create a system that is able to

238

Fathallah, A., El-Yacoubi, M. and Ben Amara, N.

EHDI: Enhancement of Historical Document Images via Generative Adversarial Network.

DOI: 10.5220/0011662700003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 4: VISAPP, pages

238-245

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

effectively enhance document images. An ideal sys-

tem should be able to remove noise and watermarks

while also maintaining the quality of the text in doc-

ument images. The ability to perform both tasks si-

multaneously would be highly desirable. Deep neural

networks, speciﬁcally deep convolutional neural net-

works (auto-encoders and variational auto-encoders

(VAE)) (Mao et al., 2016; Dong et al., 2015), and

generative adversarial networks (GANs) (Isola et al.,

2017), have recently made signiﬁcant progress in gen-

erating and restoring natural images.

The remainder of this paper is organized as fol-

lows. Section 2 provides an overview of relevant pre-

vious work. Section 3 presents the proposed approach

for enhancing degraded documents. The experimental

study is described in section 4. Finally, the conclusion

and future directions are outlined in Section in section

2 RELATED WORK

Document enhancement involves the perceptual qual-

ity improvement of document images and the removal

of degradation effects and artifacts from the images

to make them look as they originally did (Hedjam

and Cheriet, 2013). In order to improve the qual-

ity of historical documents, the binarization of docu-

ments is the most widely applied technique. It aims

to separate each pixel of text from the background

(Sauvola and Pietik

ainen, 2000). This technique re-

duces the amount of noise in document images. Tradi-

tional methods of document binarization (Otsu, 1979;

Sauvola and Pietik

ainen, 2000) are generally formu-

lated using the thresholding technique. Hence, sev-

eral approaches have evolved to determine the most

optimal thresholds for applying it as a ﬁlter. Depend-

ing on the threshold(s), binary classiﬁcation is per-

formed to specify whether the pixels are part of the

text or the degradation. In order to determine the bi-

narization threshold, the authors in (Chou et al., 2010)

proposed a method based on machine learning tech-

niques. Each region of the image is deﬁned as a three-

dimensional feature vector obtained from the gray-

level pixel value distribution. To classify each region

into one of four different threshold values, the sup-

port vector machine (SVM) was employed. Gener-

ally, the major limitation of these traditional methods

lies in the fact that the results are extremely depen-

dent on the document conditions. Indeed, in the pres-

ence of a complex image background or with an atyp-

ical intensity, several problems appear. A variational

model aimed at eliminating transparencies from de-

graded two-sided document images is introduced in

(Moghaddam and Cheriet, 2009).

Recently, GANs have gained impressive achieve-

ments in both image generation and translation. In

this section, we consider the application of GANs

in document processing and enhancement issues. In

terms of image segmentation, Ledig et al. reported

SRGAN (Ledig et al., 2017), which provides a GANs

for super-resolution of images. The authors employed

conditional GANs for document enhancement tasks

based on image-to-image translation. Dual GAN gen-

erator algorithms (Yi et al., 2017) intended for under-

water image enhancement exploit two or more gen-

erators for predicting the enhanced image. The in-

tention behind using two generators along with one

discriminator or two generators with two discrimina-

tors is to either share features between the generators

or to consider the prediction of one generator as an

input to the other generator. The model is weakly

supervised and avoids the need to use matched un-

derwater images for training in that the underwater

images can be considered in unknown locations, thus

allowing for adversarial learning. Isola et al. (Isola

et al., 2017) developed the GAN Pix2Pix for image-

to-image translation using CGAN. An adversarial loss

is used to train the GAN Pix2Pix model generator,

thus promoting plausible image generation in the tar-

get domain.The discriminator assesses whether the

generated image is a real transformation of the source

image. There are various approaches to improving

documents, but most of these approaches address a

speciﬁc problem. For example, in a study published

in (Souibgui and Kessentini, 2020), the authors ex-

amined the issue of documents that have been dam-

aged due to watermarks or stamps and proposed a

solution using conditional GANs to restore historical

documents to their original, undamaged state. Sim-

ilarly, the authors in (Jemni et al., 2022) developed

a method using GANs that combines document bi-

narization with a recognition stage. Another study

(Gangeh et al., 2021) speciﬁcally addressed problems

such as blurry text, salt and pepper noise, and water-

marks by proposing a uniﬁed architecture that com-

bines a deep network with a cycle-consistent GAN for

the purpose of denoising document images.

As motioned above, there are some recent tech-

niques for document enhancement that are emerging,

involving the use of deep learning tools. Most of them

employ Convolutional Neural Networks (CNNs) and

GANs, in order to learn how to generate a clean

binary version for any degraded document image

(Souibgui and Kessentini, 2020; Tamrin et al., 2021).

However, most authors employed a simple architec-

ture with an optimization technique that is not well

adapted to the complexity of historical document im-

EHDI: Enhancement of Historical Document Images via Generative Adversarial Network

239

ages. Essentially, the aim of document enhancement

is to generate a signiﬁcantly better and cleaner version

of degraded document images, which is very bene-

ﬁcial for further processing tasks. To this end, we

propose a robust GAN architecture that is trained and

optimized following several loss functions in order to

overcome the complexity of historical documents and

to generate a quite clean image comparing with the

existing methods.

3 PROPOSED METHOD

In this section, we present the main steps of our pro-

posed approach called EHDI (Enhancement of Histor-

ical Document Images via GANs). It aims to generate

a clear version of a degraded historical document us-

ing GANs. GANs are machine learning models that

are used to learn the distribution of real data and gen-

erate images based on random noise. The goal of

our model is to use GANs to improve the quality of

historical documents. The main purpose of GANs,

as previously discussed in studies such as (Marnissi

et al., 2021; Souibgui and Kessentini, 2020; Jemni

et al., 2022), is to learn the distribution of real data

and create an output image from random noise. In

our GAN model, the goal is to generate a clean ver-

sion of a degraded historical document. This can be

thought of as an image-to-image conversion process,

where our model learns to map a degraded document

image (represented as ”x”) to a clean document image

(represented as ”y”). During the training process, the

GAN takes a degraded document image as input and

attempts to generate a cleaned version of it. On the

other hand, the discriminator receives two inputs: the

generated image and the ground truth, which is the

known clean version of the degraded image. It then

determines whether the generated image is a realistic

representation or not based on the ground truth. As

shown in Figure 1, our model consists of two main

components: a generator (G) and a discriminator (D).

The generator is trained to convert a degraded docu-

ment image into a clean version, while the discrim-

inator helps the generator to produce more realistic

images by distinguishing between generated and real

images.

3.1 Generator Architecture

The generator in our model is an image transforma-

tion network that generates the transformed image us-

ing the input image. It is designed as an autoencoder

model, which consists of an encoder and a decoder.

The input image is typically processed through a se-

ries of convolutional layers with downward sampling

to reach a particular layer, and then decoded through

a series of up-sampling and convolutional layers. Fig-

ure 1 illustrates the details of our suggested GAN ar-

chitecture. Our generator network is based on the ar-

chitecture proposed in (Wang et al., 2018), with each

sub-network following the structure outlined in (John-

son et al., 2016).

3.2 Discriminator Architecture

The discriminator in our model is a simple fully con-

volutional network that receives two input images: the

generated image and its ground truth. Its purpose is

to determine whether the generated image is real or

fake. As shown in Figure 1, our suggested discrim-

inator architecture consists of ﬁve convolutional lay-

ers, followed by a normalization layer (except for the

last layer) and a LeakyReLU activation function (ex-

cept for the ﬁrst and last layers). Inspired by Patch-

GAN (Isola et al., 2017), we use a 70×70 patch as in-

put to our discriminative network, which determines

whether local image patches are real or fake. The dis-

criminator’s goal is to identify whether the input patch

in an image is genuine or synthetic.

3.3 Loss Functions of Proposed GAN

To effectively train our EHDI model, we include a

content loss to penalize the distance between the gen-

erated and ground-truth images. The adversarial dis-

criminator helps the generator to synthesize ﬁne and

speciﬁc details. The discriminator helps the generator

create more accurate and speciﬁc details by identify-

ing and penalizing deviations from the desired output.

To further enhance the clarity and precision of these

details, we also incorporate a combination of percep-

tual and Total Variation (TV) losses. Our objective

loss is comprised of four losses: adversarial loss, con-

tent loss, perceptual loss, and TV loss. These losses

are deﬁned as follows:

• Adversarial loss: To encourage the generator to

produce high-quality, accurate images, we use an

adversarial loss. This loss function, deﬁned in

Eq.(1), helps ensure that the generated clean im-

ages G(x) are as close as possible to the true clean

images y.

adv

= E

[−log(D(G(x), y)] (1)

• Content loss: To preserve the content informa-

tion present in the ground truth image y in the

generated image G(x), we use a content loss in

our improved GAN. This loss function, deﬁned

in Eq.(2), is a pixel-wise mean squared error that

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

240

Block 1

AvgPool 3x3

3x3x64

Generated

image

Block n

...

Block 1

Block m

Degraded

document image

Ground truth

Block

Conv

InstanceNorm

ReLU

LeakyReLU

Tanh

Sigmoid

3x3x128

3x3x256

3x3x128 3x3x256

3x3x256

7x7x32

3x3x64

3x3x32

3x3x3

4x4x32

4x4x64

4x4x128

4x4x256

4x4x1

VGG-16

7x7x64

𝐿

𝑐𝑛𝑡

𝐿

𝑇𝑉

𝐿

𝑎𝑑𝑣

𝐿

𝑝𝑟𝑐

Discriminator

Generator

Figure 1: Proposed GAN architecture for our document enhancement model EHDI.

minimizes low-level content errors between the

generated cleaned images and their corresponding

ground truth images.

content

W H

∑

i=1

∑

j=1

∥ y

i, j

− G(x)

i, j

∥

(2)

where W and H refer to the height and width of

the degraded image, respectively. The L

norm

is denoted by ∥ . ∥

, and the pixel values of the

ground truth image and the generated image are

represented by y

i, j

and G(x)

i, j

, respectively.

• Perceptual loss: To improve the perceptual quality

of the generated results and correct any distorted

textures caused by the adversarial loss, we use a

perceptual loss function introduced in (Johnson

et al., 2016). This loss function calculates the dis-

tance between the generated image and its ground

truth based on high-level representations extracted

from a pre-trained VGG-16 model.The perceptual

loss is deﬁned by Eq.(3).

prc

∑

i=1

∑

j=1

∥ Φ

(y)

i, j

−Φ

(G(x))

i, j

∥

(3)

where Φ

represents the feature representations of

the k

maxpooling layer in the VGG-16 network,

and C

represents the size of these feature

representations.

• Total variation loss: To prevent over-pixelization

and improve the spatial smoothness of the cleaned

document images, we use the TV loss introduced

in (Aly and Dubois, 2005). This loss function is

deﬁned in Eq.(4).

W H

∑

|∇

G( ˜y) + ∇

G( ˜y)| (4)

where |.| refers to the absolute value per element

of the indicated input.

The loss function that optimizes the network param-

eters of the generator (G) is referred to as the global

loss function L given by Eq.(5).

L = L

cnt

+ λ

adv

+ λ

prc

+ λ

(5)

where λ

adv

, λ

prc

and λ

represent the weights that

control the share of different losses in the full objec-

tive function.

4 EXPERIMENTS

In this section, we present the main experiments that

were conducted to evaluate the effectiveness of our

proposed approach.

4.1 Datasets

To train our document enhancement architecture, we

used the Noisy Ofﬁce database (Zamora-Mart

ınez

et al., 2007). For the evaluation phase, we con-

sidered the DIBCO 2013 (Pratikakis et al., 2013),

DIBCO 2017 (Pratikakis et al., 2017), H-DIBCO

2018 (Pratikakis et al., 2018b) and (Pantke et al.,

2014) datasets.

4.2 Experimental Setup

4.2.1 Evaluation Protocol

To evaluate the ability and the quality of our EHDI

model, we have conducted a series of experiments

on different datasets. We introduce qualitative and

quantitative results to evaluate our EHDI. We se-

lect four performance assessments (Pratikakis et al.,

2013): Peak signal-to-noise ratio (PSNR), pseudo-F-

measure (F

), Distance reciprocal distortion metric

(DRD) and F-measure.

EHDI: Enhancement of Historical Document Images via Generative Adversarial Network

241

4.2.2 Implementation Details

The learning process is optimized using the stochas-

tic gradient descent algorithm with a learning rate of

−3

and a batch size of 512. T. All parameter values

were chosen based on empirical testing. The exper-

iments were implemented using the PyTorch frame-

work and the EHDI model was trained on an NVIDIA

Quadro RTX 6000 GPU with 24 GB of RAM. To fa-

cilitate the training of the architecture, each image is

resized to 1024 × 1024 pixels and a set of stacked

patches of size 256 × 256 pixels are extracted. This

results in a total of 2,304 patch pairs, which are used

to train the EHDI model. During training, the weights

of the different losses in the full objective function are

set to λ

adv

= 0.3, λ

prc

= 1, and λ

= 1, respectively.

4.3 Results

In this section, we evaluate the performance of the

proposed EHDI model and compare it to the cur-

rent state of the art in document binarization. It is

important to note that the model was only trained

on the Noisy Ofﬁce (Zamora-Mart

ınez et al., 2007)

database, while the evaluation was conducted on ex-

ternal databases not included in the training. The re-

sults of our EHDI model are presented in Table 1 and

compared to other approaches on the DIBCO 2013

dataset. Figure 2 also shows a qualitative comparison

of the results on the DIBCO 2013 dataset, where it can

be seen that the EHDI model produces a cleaner im-

age quality than DE-GAN, especially when the degra-

dation is very dense. This is because DE-GAN may

struggle to remove such degradation from the docu-

ment background in these cases.

Table 1: Results of our proposed EHDI on DIBCO 2013

dataset.

Model PSNR F-measure F

DRD

(Otsu, 1979) 16.6 83.9 86.5 11.0

(Niblack, 1985) 13.6 72.8 72.2 13.6

(Sauvola and Pietik

ainen,

2000)

16.9 85.0 89.8 7.6

(Gatos et al., 2004) 17.1 83.4 87.0 9.5

(Su et al., 2012) 19.6 87.7 88.3 4.2

(Tensmeyer and Martinez,

2017)

20.7 93.1 96.8 2.2

(Xiong et al., 2018) 21.3 93.5 94.4 2.7

(Vo et al., 2018) 21.4 94.4 96.0 1.8

(Howe, 2013) 21.3 91.3 91.7 3.2

(Souibgui and Kessentini,

2020)

24.9 99.5 99.7 1.1

EHDI 26.8 99.9 99.9 0.97

To demonstrate the improvement achieved by our

EHDI model, Figure 3 presents an example of the

generated document image that is very close to and

even superior to the ground truth image.

Original image

Ground truth

DE-GAN (Souibgui and Kessentini, 2020)

EHDI

Figure 2: Example of degraded documents enhancement by

our EHDI and DE-GAN on sample PR08 from DIBCO-

2013 (Souibgui and Kessentini, 2020).

Original image Ground truth Generated image

Figure 3: Example of enhancing degraded documents by

our EHDI.

The proposed EHDI (enhanced document im-

age) outperforms the state-of-the-art DE-GAN model

(Souibgui and Kessentini, 2020). This is evident in

the results of the 2017 DIBCO and 2018 H-DIBCO

test sets, as shown in Table 2. An example of this

superiority is shown in Figure 4, where EHDI outper-

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

242

Table 2: A comparative review of competitor approaches of DIBCO 2018 on DIBCO 2017 and DIBCO 2018 Datasets.

Model

DIBCO 2018 DIBCO2017

PSNR F-measure F

DRD PSNR F-measure F

DRD

1 (Pratikakis et al., 2018a) 19.11 88.34 90.24 4.92 17.99 89.37 90.17 5.51

7 (Pratikakis et al., 2018a) 14.62 73.45 75.94 26.24 15.72 84.36 87.34 7.56

2 (Pratikakis et al., 2018a) 13.58 70.04 74.68 17.45 14.04 79.41 82.62 10.70

3b (Pratikakis et al., 2018a) 13.57 64.52 68.29 16.67 15.28 82.43 86.74 6.97

6 (Pratikakis et al., 2018a) 11.79 46.35 51.39 24.56 15.38 80.75 87.24 6.22

(Souibgui and Kessentini, 2020) 16.16 77.59 85.74 7.93 18.74 97.91 98.23 3.01

EHDI 20.31 92.69 90.83 3.94 19.15 98.56 99.44 2.87

Original image Ground truth Winner team DE-GAN EHDI

Figure 4: Qualitative binarization results on sample 16 in DIBCO 2017 dataset. Here, we compare the results of our proposed

model with the winner’s approach (Pratikakis et al., 2017) and DE-GAN (Souibgui and Kessentini, 2020).

Original image DE-GAN (Souibgui and Kessentini, 2020) EHDI

Figure 5: Example of qualitative results on HADARA dataset compared to DE-GAN enhancement approach (Souibgui and

Kessentini, 2020).

forms the winner’s method and DE-GAN on (sample

16) from DIBCO 2017. The winner’s method used a

U-net architecture and data augmentation techniques,

while DE-GAN used a simple GAN network. Our

model achieved better results due to the use of multi-

ple loss functions that optimize the generator to pro-

duce images more closely aligned with the ground

truth.

As shown in Figure 5, the visual quality of the

enhanced document images on samples from the

HADARA dataset (Pantke et al., 2014) is demon-

strated. It is clear that EHDI consistently outperforms

DE-GAN. To fairly evaluate our EHDI against state-

of-the-art methods, we used the same sample as in the

DE-GAN paper and compared EHDI to pix2pix-HD

(Wang et al., 2018) and CycleGan (Zhu et al., 2017).

The results, shown in Figure 6 and Table 3, demon-

strate the superior performance of EHDI in terms

of visual quality compared to cycleGAN, pix2pix-

HD, and DE-GAN. In contrast to previous work in

(Souibgui and Kessentini, 2020), which used three

separate cGAN-based models for binarization, water-

EHDI: Enhancement of Historical Document Images via Generative Adversarial Network

243

marking, and deblurring, this paper presents a single

model that can handle all these tasks.

Table 3: Results of our proposed model on DIBCO 2018

dataset.

Model PSNR F-measure F

DRD

cycleGAN 11.00 56.33 58.07 30.07

pix2pix-HD 14.42 72.79 76.28 15.13

DE-GAN 16.16 77.59 85.74 7.93

EHDI 26.8 99.9 99.9 0.97

Original image

Ground truth

CycleGan

Pix2pix-HD

DE-GAN (Souibgui and Kessentini, 2020)

EHDI

Figure 6: Example of qualitative enhancement results pro-

duced by different models of the sample (9) from the H-

DIBCO 2018 dataset.

5 CONCLUSION

In this paper, we have put forward a conditional GAN

as a means to generate clean document images from

highly degraded images. Our suggested EHDI has

been designed to handle different degradation tasks

such as watermark removal and chemical degradation

with the goal of producing hyper-clean document im-

ages and ﬁne detail recovery performances.

Extensive experiments have shown the effective-

ness of the proposed EHDI for cleaning extremely de-

graded documents. An interesting improvement for

historical documents compared to many recent state-

of-the-art methods on reference datasets.

Future work will include the adoption of the vi-

sion transformer techniques for a better document im-

provement process. In addition, we intend to add

a recognition module to our framework to provide

a comprehensive platform for processing historical

documents.

REFERENCES

Aly, H. A. and Dubois, E. (2005). Image up-sampling us-

ing total-variation regularization with a new observa-

tion model. IEEE Transactions on Image Processing,

14(10):1647–1659.

Chen, X., He, X., Yang, J., and Wu, Q. (2011). An effective

document image deblurring algorithm. In CVPR 2011,

pages 369–376. IEEE.

Chou, C.-H., Lin, W.-H., and Chang, F. (2010). A bina-

rization method with learning-built rules for document

images produced by cameras. Pattern Recognition,

43(4):1518–1530.

Dong, C., Loy, C. C., He, K., and Tang, X. (2015). Image

super-resolution using deep convolutional networks.

IEEE transactions on pattern analysis and machine

intelligence, 38(2):295–307.

Finlayson, G. D., Hordley, S. D., and Drew, M. S. (2002).

Removing shadows from images. In European con-

ference on computer vision, pages 823–836. Springer.

Gangeh, M. J., Plata, M., Motahari, H., and Duffy, N. P.

(2021). End-to-end unsupervised document image

blind denoising. arXiv preprint arXiv:2105.09437.

Gatos, B., Pratikakis, I., and Perantonis, S. J. (2004). An

adaptive binarization technique for low quality histor-

ical documents. In International Workshop on Docu-

ment Analysis Systems, pages 102–113. Springer.

Hedjam, R. and Cheriet, M. (2013). Historical document

image restoration using multispectral imaging system.

Pattern Recognition, 46(8):2297–2312.

Howe, N. R. (2013). Document binarization with automatic

parameter tuning. International journal on document

analysis and recognition (ijdar), 16(3):247–258.

Isola, P., Zhu, J.-Y., Zhou, T., and Efros, A. A. (2017).

Image-to-image translation with conditional adversar-

ial networks. In Proceedings of the IEEE conference

on computer vision and pattern recognition, pages

1125–1134.

Jemni, S. K., Souibgui, M. A., Kessentini, Y., and Forn

es,

A. (2022). Enhance to read better: A multi-task ad-

versarial network for handwritten document image en-

hancement. Pattern Recognition, 123:108370.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

244

Johnson, J., Alahi, A., and Fei-Fei, L. (2016). Perceptual

losses for real-time style transfer and super-resolution.

In European conference on computer vision, pages

694–711. Springer.

Ledig, C., Theis, L., Husz

ar, F., Caballero, J., Cunningham,

A., Acosta, A., Aitken, A., Tejani, A., Totz, J., Wang,

Z., et al. (2017). Photo-realistic single image super-

resolution using a generative adversarial network. In

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition, pages 4681–4690.

Mao, X., Shen, C., and Yang, Y.-B. (2016). Image restora-

tion using very deep convolutional encoder-decoder

networks with symmetric skip connections. Advances

in neural information processing systems, 29.

Marnissi, M. A., Fradi, H., Sahbani, A., and Essoukri

Ben Amara, N. (2021). Thermal image enhancement

using generative adversarial network for pedestrian

detection. In 2020 25th International Conference on

Pattern Recognition (ICPR), pages 6509–6516. IEEE.

Moghaddam, R. F. and Cheriet, M. (2009). A variational

approach to degraded document enhancement. IEEE

transactions on pattern analysis and machine intelli-

gence, 32(8):1347–1361.

Niblack, W. (1985). An introduction to digital image pro-

cessing. Strandberg Publishing Company.

Otsu, N. (1979). A threshold selection method from gray-

level histograms. IEEE transactions on systems, man,

and cybernetics, 9(1):62–66.

Pantke, W., Dennhardt, M., Fecker, D., M

argner, V., and

Fingscheidt, T. (2014). An historical handwritten

arabic dataset for segmentation-free word spotting-

hadara80p. In 2014 14th International Conference on

Frontiers in Handwriting Recognition, pages 15–20.

IEEE.

Pratikakis, I., Gatos, B., and Ntirogiannis, K. (2013). Ic-

dar 2013 document image binarization contest (dibco

2013). In 2013 12th International Conference on Doc-

ument Analysis and Recognition, pages 1471–1476.

IEEE.

Pratikakis, I., Zagoris, K., Barlas, G., and Gatos, B. (2017).

Icdar2017 competition on document image binariza-

tion (dibco 2017). In 2017 14th IAPR International

Conference on Document Analysis and Recognition

(ICDAR), volume 1, pages 1395–1403. IEEE.

Pratikakis, I., Zagoris, K., Kaddas, P., and Gatos, B.

(2018a). Icfhr 2018 competition on handwritten doc-

ument image binarization (h-dibco 2018). 2018 16th

International Conference on Frontiers in Handwriting

Recognition (ICFHR), pages 489–493.

Pratikakis, I., Zagoris, K., Kaddas, P., and Gatos, B.

(2018b). Icfhr2018 competition on handwritten doc-

ument image binarization contest (h-dibco 2018). In

International conference on frontiers in handwriting

recognition (ICFHR). IEEE, pages 1–1.

Sauvola, J. and Pietik

ainen, M. (2000). Adaptive document

image binarization. Pattern recognition, 33(2):225–

236.

Souibgui, M. A. and Kessentini, Y. (2020). De-gan: A con-

ditional generative adversarial network for document

enhancement. IEEE Transactions on Pattern Analysis

and Machine Intelligence.

Su, B., Lu, S., and Tan, C. L. (2012). Robust docu-

ment image binarization technique for degraded docu-

ment images. IEEE transactions on image processing,

22(4):1408–1417.

Tamrin, M. O., El-Amine Ech-Cherif, M., and Cheriet,

M. (2021). A two-stage unsupervised deep learning

framework for degradation removal in ancient docu-

ments. In International Conference on Pattern Recog-

nition, pages 292–303. Springer.

Tensmeyer, C. and Martinez, T. (2017). Document image

binarization with fully convolutional neural networks.

In 2017 14th IAPR international conference on doc-

ument analysis and recognition (ICDAR), volume 1,

pages 99–104. IEEE.

Vo, Q. N., Kim, S. H., Yang, H. J., and Lee, G. (2018).

Binarization of degraded document images based on

hierarchical deep supervised network. Pattern Recog-

nition, 74:568–586.

Wang, T.-C., Liu, M.-Y., Zhu, J.-Y., Tao, A., Kautz, J., and

Catanzaro, B. (2018). High-resolution image synthe-

sis and semantic manipulation with conditional gans.

In Proceedings of the IEEE conference on computer

vision and pattern recognition, pages 8798–8807.

Xiong, W., Jia, X., Xu, J., Xiong, Z., Liu, M., and Wang,

J. (2018). Historical document image binarization

using background estimation and energy minimiza-

tion. In 2018 24th International Conference on Pat-

tern Recognition (ICPR), pages 3716–3721. IEEE.

Yi, Z., Zhang, H., Tan, P., and Gong, M. (2017). Dual-

gan: Unsupervised dual learning for image-to-image

translation. In Proceedings of the IEEE international

conference on computer vision, pages 2849–2857.

Zamora-Mart

ınez, F., Espa

na-Boquera, S., and Castro-

Bleda, M. (2007). Behaviour-based clustering of neu-

ral networks applied to document enhancement. In In-

ternational Work-Conference on Artiﬁcial Neural Net-

works, pages 144–151. Springer.

Zhu, J.-Y., Park, T., Isola, P., and Efros, A. A. (2017).

Unpaired image-to-image translation using cycle-

consistent adversarial networks. In Proceedings of

the IEEE international conference on computer vi-

sion, pages 2223–2232.

EHDI: Enhancement of Historical Document Images via Generative Adversarial Network

245