Enhancing the Readability of Palimpsests Using Generative Image

Inpainting

Mahdi Jampour

, Hussein Mohammed

and Jost Gippert

Cluster of Excellence, Understanding Written Artefacts, Universit

at Hamburg, Hamburg, Germany

Keywords:

Generative AI, Inpainting, Deep CNN, Historical Manuscripts, Image Enhancement, Palimpsests.

Abstract:

Palimpsests are manuscripts that have been scraped or washed for reuse, usually as another document. Recov-

ering the undertext of these manuscripts can be of signiﬁcant interest to scholars in the humanities. Multispec-

tral imaging is a technique often used to make the undertext visible in palimpsests. Nevertheless, this approach

is not sufﬁcient in many cases, due to the fact that the undertext in resulting images is still covered by the over-

text or other artefacts. Therefore, we propose deﬁning this issue as an inpainting problem and enhancing

the readability of the undertext using generative image inpainting. To this end, we introduce a novel method

for generating synthetic multispectral palimpsest images and make the generated dataset publicly available.

Furthermore, we utilise this dataset in the ﬁne-tuning of a generative inpainting approach to enhance the read-

ability of palimpsest undertext. The evaluation of our approach is provided for both the synthetic dataset and

palimpsests from actual research in the humanities. The evaluation results indicate the effectiveness of our

method in terms of both quantitative and qualitative measures.

1 INTRODUCTION

Palimpsests are documents or writing substrates

where new text has been written over previously

erased content. Historical constraints, including the

cost and complexity of production, led to repurpos-

ing manuscripts by removing old content through

mechanical or chemical methods. The erased text,

termed undertext, is replaced with new handwriting,

known as overtext. Scholars in the humanities ﬁnd

signiﬁcant interest in retrieving the undertext from

palimpsests.

Revealing the undertext in palimpsest manuscripts

poses challenges due to degradation, occlusion by

overtext, and other visual artefacts. Multispectral

imaging (MSI) is a common technique, capturing

light within speciﬁc wavelength ranges to make the

undertext visible. While MSI aids visibility, it may

not render the text readable in many cases. Figure 1

illustrates the difference between MSI and enhanced

readability for a palimpsest document. Parts of the

undertext obscured by overtext and artifacts cannot

be fully revealed by MSI alone. To address this, we

https://orcid.org/0000-0002-1559-1865

https://orcid.org/0000-0001-5020-3592

https://orcid.org/0000-0002-2954-340X

Figure 1: An example of a palimpsest image. (left) Origi-

nal document, (middle) Multispectral image, and (right) En-

hanced undertext.

propose employing a generative inpainting approach

to generate these occluded pixels, guided by informa-

tion about the shape of letters in a particular script.

Efforts to separate undertext from overtext based on

aging differences and visual contrast using statistical

methods have been made. However, these methods

are effective only when clear distinctions exist be-

tween undertext and overtext. In contrast, this study

employs deep learning to reconstruct separated under-

text using models trained on the same letterforms. An

end-to-end approach is proposed to reconstruct the

undertext in multispectral palimpsest images within

its original visual context. Deﬁning this task as an

image inpainting problem, we treat overtext and vi-

Jampour, M., Mohammed, H. and Gippert, J.

Enhancing the Readability of Palimpsests Using Generative Image Inpainting.

DOI: 10.5220/0012347100003654

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

In Proceedings of the 13th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2024), pages 687-694

ISBN: 978-989-758-684-2; ISSN: 2184-4313

687

sually similar artefacts as masked regions, aiming to

reconstruct the underlying undertext portions.

Image inpainting involves reconstructing regions

within an image by ﬁlling in areas that need modiﬁ-

cation. This process is considered ill-posed because

there’s no single deﬁnitive solution for the pixel val-

ues in the target regions. Image inpainting is essen-

tial in applications such as image/video restoration,

removal, and etc.

We reframe the problem as an image restoration

task, treating overtext and visually similar artefacts

as concealed areas, with the objective of reconstruct-

ing the undertext beneath them. To achieve this, we

employ the MAT model (Li et al., 2022), a state-of-

the-art transformer-based approach, as an end-to-end

solution to enhance palimpsest manuscript readabil-

ity. Additionally, we introduce a method to generate

a training dataset of multispectral palimpsest images

for ﬁne-tuning. Our third contribution is the provi-

sion of a dataset containing 1000 synthetic samples

of multispectral Georgian palimpsest images, each

paired with a corresponding mask and ground-truth,

resulting in 3000 images. The Synthetic Georgian

Palimpsest (SGP) images dataset is publicly available

for research purposes in our research data repository

(Jampour et al., 2023a).

2 RELATED WORK

Most prior methods to improve the legibility of un-

derlying text in palimpsest multispectral images have

concentrated on segregating undertext from overtext.

In contrast, our approach aims to offer an end-to-end

solution, focusing on the complete reconstruction of

undertext within the original visual context of mul-

tispectral palimpsest images. This objective is cast

as an image inpainting challenge, where overtext and

visually analogous artefacts are regarded as masked

regions, and the occluded undertext undergoes recon-

struction.

2.1 Readability Enhancement of

Palimpsests

Numerous studies have tackled the challenge of

improving the readability of underlying text in

palimpsests, given their signiﬁcance in manuscript

research. Early works, such as that of Salerno

et al. (Salerno et al., 2007), introduced statistical

methodologies, notably Principal Component Anal-

ysis (PCA) and Independent Component Analysis

(ICA), to separate undertext from overwritten text in

palimpsest multispectral images. These techniques

aimed to enhance the legibility of concealed text in

such images. It is crucial to note, however, that PCA,

while effective for noise reduction and data simpliﬁ-

cation, operates under the assumption of linear rela-

tionships between variables, a limitation that may not

align with the inherent complexities of palimpsest en-

hancement challenges.

In a different study, Easton et al. (Easton et al.,

2011) applied classical optical imaging and well-

established image processing techniques to enhance

the readability of the Archimedes Palimpsest. Sim-

ilar methods, grounded in traditional image process-

ing, signiﬁcantly contributed to the decipherment of

obscured text.

More recently, Starynska et al. (Starynska et al.,

2021) leveraged deep generative networks to sepa-

rate undertext from overtext in palimpsests. This

approach necessitated a signiﬁcant volume of data

similar to the typeface of the undertext, which they

generated synthetically. The authors used the gen-

erative model to separate undertext and restore the

original text. On the other hand, a recent study by

Obukhova et al. (Obukhova et al., 2023) rightly un-

derscored the data requirements associated with deep

learning approaches. Recognizing the need for ex-

tensive data, they evaluated the legibility of undertext

in palimpsest images using several classical statisti-

cal methods, including PCA and Linear Discriminant

Analysis (LDA).

2.2 Image Inpainting

Traditional image inpainting methods, including

exemplar-based techniques, typically rely on patch-

based algorithms that primarily utilize local informa-

tion around the target region. This approach has limi-

tations when tasked with inpainting extensive image

regions or when considering global visual features,

such as a dominant background texture. Both tradi-

tional and deep-learning-based approaches face chal-

lenges posed by the visual complexity of images (Xi-

ang et al., 2023). In the dataset used for this study, an

additional signiﬁcant challenge arises due to the ne-

cessity of achieving a very high level of ﬁdelity. This

is primarily attributed to the fact that even slight ad-

ditions, removals, or alterations in the orientation of

a few pixels can result in entirely incorrect charac-

ters within the underlying text. Moreover, the target

regions constitute a substantial portion of the image,

yet are distributed across a large number of small lo-

cal sections.

In works, Bertalmio et al. (Bertalmio et al., 2000)

introduced the concept of utilizing information from

the surroundings of inpainting regions. Subsequently,

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

688

(a) (b) (c) (d)

Figure 2: Synthesizing MSI images of palimpsests. (a) Generated background, (b) rendered undertext on generated back-

ground, (c)Adding randomly generated overtext, (d) Adding randomly generated noise to mimic the typical texture in MSI

images. The second row shows zoomed regions in the images.

enhanced variations of Bertalmio’s approach were de-

veloped, known as exemplar-based image inpaint-

ing (Abdulla and Ahmed, 2021; Zhang et al., 2019),

Markov Random Field (MRF) techniques (Ru

c and

zurica, 2015; Jampour et al., 2017), which rely on

surrounding information for pixel value estimation. In

certain scenarios, high-level attributes are harnessed

to improve inpainting results. For example, (Jam-

pour et al., 2017) proposed employing a guided face

with MRF for inpainting facial features. However, the

challenge of inpainting large regions presents a signif-

icant hurdle, where deep-learning-based approaches

yield more favourable results. For instance, MAT (Li

et al., 2022) stands out as one of the state-of-the-art

method for large-hole image inpainting, harnessing

transformers to handle high-resolution images and ef-

fectively integrate local and global information. Simi-

larly, recent deep-learning approaches, including gen-

erative models (e.g., GANs), effectively address the

challenge of missing region size, achieving notable

success in image inpainting.

3 PROPOSED APPROACH

This section will commence with a presentation of

the problem statement. Subsequently, we will out-

line the process of generating synthetic MSI images

of palimpsests and introduce the Synthetic MSI Im-

ages of Georgian Palimpsests (SGP dataset). Follow-

ing this, we will describe the process of selecting a

pre-training dataset for the MAT model and elaborate

on the ﬁne-tuning of the pre-trained MAT model.

3.1 Problem Statement

In contrast to Starynska et al.’s approach (Starynska

et al., 2021), which involves a separating undertext

model, and statistical models like Obukhova et al.’s

(Obukhova et al., 2023), we propose to reconstruct the

undertext in multispectral palimpsest images within

its original visual context. We frame this task as an

image inpainting problem, where overtext and visu-

ally similar artefacts are considered masked regions,

and the underlying undertext portions need recon-

struction. Let P

image

denote the set of pixel values

in a palimpsest multispectral image, including back-

ground, undertext, overwritten text, and noise. A pos-

sible formulation of the problem can be expressed as:

image

= Ψ(P

target

⊕ P

noise

)

(1)

Where P

target

and P

noise

are two sets of pixel values in

the image representing the undertext and the overtext

with the noise correspondingly, and both are subject

to the Ψ(.) function representing general background

texture of MSI images. We deﬁne symbol ⊕ to show

that these two sets (i.e., P

target

and P

noise

) are mixed

non-linearly, and their intersection is non-empty, in

another word P

target

∩ P

noise

/∈

0. Therefore, there is

a subset of pixel values J which belongs to both sets

target

and P

noise

where:

J ⊂ P

target

, and J ⊂ P

noise

J = P

target

∩ P

noise

(2)

Our objective is to estimate J, and the primary

challenge arises from the absence of information re-

garding the pixel values of the undertext (i.e., P

target

Enhancing the Readability of Palimpsests Using Generative Image Inpainting

689

Figure 3: A generated sample from the proposed dataset. Left to right: (a) Ground-truth, (b) Mask, and (c) Target image.

As an inverse problem, we suggest generating a set

of pixel values in P

target

and P

noise

that are very sim-

ilar and emulating the Ψ(.) function. This approach

enables the model to learn the estimation of the over-

lapping regions (i.e., J) and subsequently perform in-

painting on the palimpsests to reconstruct the missing

portions of the undertext.

3.2 Generating Synthesised MSI Images

of Palimpsests

Generative impainting approaches, such as GANs,

yield state-of-the-art performance in various aspects

(Jampour et al., 2023b; Zare et al., 2019). Neverthe-

less, they require a large amount of training data with

ground-truth images and pixel-level annotated masks.

Creating this kind of training data is not only time

consuming, but also not possible due to the unavail-

ability of such information. Therefore, we propose to

generate synthetic MSI images of palimpsests. To this

end, we create images in four steps as follows:

1. Generating synthesised background-texture of

MSI images (i.e., Ψ(.)).

2. Rendering undertext using letters with similar

typeface and mixing the rendered text with the

generated background (i.e., Ψ(P

target

)).

3. Generating random overtext and mixing it with

the resulting image from the previous step (i.e.,

Ψ(P

target

⊕ P

noise

)).

4. Adding randomly generated noise to mimic the

typical texture in MSI images.

For the generation of MSI-like background in the

training images, regions without text were cropped

from actual MSI images of palimpsests, and well-

known post-processing steps, including stitching,

scaling, and cleaning, were applied. Figure 2a illus-

trates an example of the background preparation. The

undertext is rendered using a typeface that was cre-

ated from these palimpsest images. Figure 2b show-

cases an example of the resulting images.

The MSI images used pertain to Georgian

palimpsests from the 11

century, with undertext in

the Caucasian Albanian language from the 7

-8

centuries, stored in the library of St. Catherine’s

Monastery on Mount Sinai (Sin. georg. NF 13 and

55). The source of this MSI data is furnished by the

Sinai Palimpsests Project via the Sinai Manuscripts

Digital Library at UCLA. The initial draft of the type-

face was crafted by Jost Gippert in 2005, and the ﬁ-

nal version was prepared by Andreas St

otzner in 2007

(Gippert and Dum-Tragut, 2023).

Subsequently, random overtext is generated and

added to our synthetic images (i.e., Ψ(P

target

⊕

noise

)). Figure 2c presents an example of the re-

sulting images. In these images, the content of the

overtext is randomly selected, but the direction, line

spacing, and scale are made as similar as possible

to the overtext found in the actual palimpsest im-

ages. Finally, additional noise is introduced to repli-

cate the random artifacts found in the MSI images

of palimpsests, as depicted in Figure 2d. We note

that the amount, distribution, and location of noise are

considered entirely random.

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

690

(a) (b) (c) (d) (e)

Figure 4: Visual comparison between the generated images using the MAT model pretrained on different datasets without

further ﬁne-tuning on the SGP dataset: (a) the synthetic input sample. The result of MAT pretrained on (b) CelebA-HQ, (c)

FFHQ, (d) Places365, (e) Places365˙FullData.

3.3 The Synthetic MSI Images of

Georgian Palimpsests (SGP)

A dataset comprising 1000 synthetic MSI images of

Georgian palimpsests has been created through the

previously outlined steps. Each sample in the dataset

is accompanied by a mask and a ground-truth image.

The ground-truth image, P

target

, serves as the target

image for evaluating the generated image using the

inpainting model. The mask image, P

noise

, delineates

the regions that require inpainting, encompassing the

overtext and other noise artifacts. Figure 3 (a)-(c) dis-

plays a synthetic sample, including its corresponding

mask and ground-truth images. All images are of uni-

form dimensions, measuring 2800 × 2100 pixels, and

are stored in colour PNG format. Consequently, this

dataset consists of a total of 3000 images across the

1000 samples and is made publicly available for re-

search purposes at (Jampour et al., 2023a).

3.4 Selection of Pre-Training Datasets

for the MAT Model

The Mask-Aware Transformer (MAT) model under-

goes pre-training using various standard and publicly

available datasets, including CelebA-HQ, FFHQ,

Places365, and an extensive version referred to as

Places365˙FullData. Initiating training from scratch

on these large datasets can be time-consuming and

computationally intensive. Therefore, the common

practice is to use pre-trained models and ﬁne-tune

them for speciﬁc tasks. Our qualitative and quantita-

tive assessments reveal that a MAT model pre-trained

on the Places365 dataset produces the most promis-

ing results, as illustrated in Figure 4 and Table 1, re-

spectively. In this quantitative evaluation, we employ

the well-established Fr

echet Inception Distance (FID)

(Heusel et al., 2017) and the Structural Similarity In-

dex Measure (SSIM) as our chosen metrics.

The FID for multivariate normal distribution is

calculated as follows:

FID =



org

− µ

inp



− Tr(

∑

org

∑

inp

− 2

∑

org

∑

inp

)

(3)

Where I

org

and I

inp

are feature vectors of ground truth

and inpainted images. µ

org

and µ

inp

are magnitude of

the vectors I

org

and I

inp

and Tr(.),

∑

org

, and

∑

inp

are

respectively the trace and covariance matrix of vec-

tors.

The SSIM on the other hand can be described as

follows:

SSIM =

(2µ

org

inp

+ c

)(2µ

org

inp

+ c

)

(µ

org

inp

+ c

)(µ

org

inp

+ c

)

(4)

Where µ

org

, µ

inp

and σ

org

, σ

inp

are mean and

variance of feature vectors I

org

and I

inp

and c

, c

are

variables to stabilize the division.

The choice to employ the Fr

echet Inception Dis-

tance (FID), in addition to it being a widely accepted

evaluation metric, is motivated by the use of genera-

tive adversarial networks (GANs) in the MAT model.

Research, as outlined by Lucic et al. (Lucic et al.,

2018), has demonstrated that FID serves as an appro-

priate metric for quantifying the realism and diversity

of images generated by GANs. FID measures differ-

ences in the representations of features, encompassing

elements such as edges, lines, and higher-order phe-

nomena.

In contrast, the Structural Similarity Index Mea-

sure (SSIM) is a widely-used metric approximating

human visual perception of similarity. Three ran-

dom SGP dataset samples underwent inpainting using

the MAT model pre-trained on four distinct datasets.

Figure 4 visually compares the results, while Table

1 provides a quantitative analysis. The MAT model

Enhancing the Readability of Palimpsests Using Generative Image Inpainting

691

Table 1: Quantitative comparison on various pre-trained models executed on three synthetic documents before and after

inpainting with FID and SSIM metrics.

FID ↓ SSIM ↑

Test Image Pre-trainedModel Before After Before After

image 0920 CelebA-HQ 298 225 0.718 0.805

FFHQ 298 262 0.718 0.781

Places365 298 160 0.718 0.816

Places365˙FullData 298 297 0.718 0.792

image 0930 CelebA-HQ 292 194 0.591 0.747

FFHQ 292 290 0.591 0.709

Places365 292 131 0.591 0.761

Places365˙FullData 292 362 0.591 0.723

image 0940 CelebA-HQ 310 239 0.713 0.792

FFHQ 310 362 0.713 0.772

Places365 310 206 0.713 0.807

Places365˙FullData 310 400 0.713 0.784

Table 2: The inpainting results of the SSIM metric on three

randomly selected samples using the ﬁne-tuned (FT) model

on the SGP Dataset.

SSIM ↑ image 0920 image 0930 image 0940

Model Before After Before After Before After

No FT 0.718 0.816 0.591 0.761 0.713 0.807

FT 0.718 0.818 0.591 0.765 0.713 0.809

FT 0.718 0.821 0.591 0.767 0.713 0.812

FT 0.718 0.823 0.591 0.766 0.713 0.813

FT 0.718 0.823 0.591 0.764 0.713 0.814

FT 0.718 0.823 0.591 0.761 0.713 0.813

FT 0.718 0.822 0.591 0.758 0.713 0.813

trained on Places365 outperforms others in both FID

and SSIM measurements, as evident in Figure 4. Con-

sequently, the MAT model pre-trained on Places365

will be ﬁne-tuned on the SGP Dataset.

3.5 Fine-Tuning the Pre-Trained Model

The MAT model, initially pre-trained on Places365,

underwent ﬁne-tuning using the SGP Dataset to en-

hance undertext reconstruction quality. The dataset’s

1000 samples were divided into subsets: 800 for train-

ing, 100 for validation, and 100 for testing. Given the

MAT model’s 512 × 512 input size, patches of this

size were cropped from all subset images. Training

utilized 50,400 patches, while each validation and test

subset employed 6,300 patches.

Testing experiments involved dividing each input

image into 512 × 512 patches, later merged after in-

painting for image reconstruction. Metrics, includ-

ing SSIM and FID, were assessed based on the recon-

structed image. Fine-tuning spanned ten epochs until

validation loss plateaued, with no substantial perfor-

mance gain. An initial learning rate of 1e

−4

prevented

drastic changes in pre-trained model weights.

Table 3: The inpainting results using the FID metric on

three randomly selected samples using the ﬁne-tuned (FT)

model on the SGP Dataset.

FID ↓ image 0920 image 0930 image 0940

Model Before After Before After Before After

No FT 298 160 292 131 310 206

FT 298 164 292 135 310 211

FT 298 171 292 142 310 228

FT 298 179 292 145 310 244

FT 298 185 292 149 310 255

FT 298 191 292 151 310 259

FT 298 192 292 152 310 258

4 EXPERIMENTAL RESULTS

This section encompasses a quantitative evaluation of

the MAT model, pre-trained on the Places365 dataset

and ﬁne-tuned on the SGP dataset. Furthermore, it

includes a qualitative evaluation on images from the

synthetic data within the SGP dataset as well as im-

ages from real palimpsests.

4.1 Quantitative Evaluation

The SSIM and FID results appear in Tables 2 and 3,

respectively, using three randomly selected samples.

FID computation (Seitzer, 2020) necessitates multi-

Table 4: Evaluation of the test subset containing 100 sam-

ples before and after inpainting with 4

FT.

FID ↓ SSIM↑

Before After Before After

Mean 293.43 194.74 0.6952 0.8114

SD 26.58 28.81 0.049 0.028

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

692

Figure 5: Real and synthetic palimpsest documents before and after inpainting. It can be seen that the readability of undertext

has been clearly enhanced. The ﬁrst and third (Zoomed) rows are the input images before inpainting, whilst the second and

fourth rows are the output images after inpainting. The two right-most samples are real palimpsests documents along with

their inpainting results.

ple input images. We extracted 368 patches per input

image, cropped conventionally with a stride of 100

pixels and a consistent size of 512 × 512. This re-

sulted in two folders for each ground truth and target

image pair, each containing 368 patches, to compute

FID. The FID results lack a discernible improvement,

possibly due to added edges and lines affecting val-

ues. Conversely, Table 2 distinctly shows SSIM en-

hancements for all images using the ﬁne-tuned model

on the proposed SGP Dataset. The ﬁne-tuned model

is also assessed on the test dataset for future compar-

isons. Table 4 presents average values and standard

deviations for FID and SSIM measures across all test

samples in the SGP test set.

4.2 Qualitative Evaluation

The primary aim is to improve the readability of un-

dertext in palimpsest images. Visual evaluation is cru-

cial for validating this approach. The inpainting re-

sults, shown in the second row of Figure 5, demon-

strate a substantial enhancement in undertext read-

ability for both synthetic and real palimpsests. No-

tably, the reconstruction accurately restores large por-

tions of occluded undertext letters, as evident in the

zoomed regions in the two lower rows of Figure 5.

Enhancing the Readability of Palimpsests Using Generative Image Inpainting

693

5 CONCLUSIONS

Palimpsests, historical manuscripts repurposed

through scraping or washing, hold signiﬁcant value

for humanities scholars. Scholars commonly use

multispectral imaging to reveal obscured undertext

in these manuscripts. However, this method often

falls short in making the undertext readable due to

occlusion by overtext and various artefacts. This

paper proposes a novel approach to enhance undertext

readability in palimpsests through generative image

inpainting. We introduced a method to generate

the Synthetic MSI Images of Georgian Palimpsests

(SGP) dataset, publicly available. Additionally, we

ﬁne-tuned a pre-trained Mask-Aware Transformer

(MAT) model using this dataset to improve its

performance. The resulting model was quantitatively

evaluated using FID and SSIM metrics on both

synthetic SGP dataset images and real palimpsest

images. Qualitative evaluation illustrated the prac-

ticality of our approach in enhancing undertext

readability for manuscript research. The results

clearly demonstrate the effectiveness of our approach

in both quantitative and qualitative measures.

ACKNOWLEDGEMENTS

The research for this work was funded by the

Deutsche Forschungsgemeinschaft (DFG, German

Research Foundation) under Germany’s Excellence

Strategy – EXC 2176 ‘Understanding Written Arte-

facts: Material, Interaction and Transmission in

Manuscript Cultures’, project no. 390893796. The

research was conducted within the scope of the Cen-

tre for the Study of Manuscript Cultures (CSMC) at

Universit

at Hamburg.

In addition, we thank St Catherine’s Monastery

on Mt Sinai, the Early Manuscripts Electronic Li-

brary and the members of the Sinai Palimpsest Project

(http://sinaipalimpsests.org/), esp. Michael Phelps,

Claudia Rapp and Keith Knox, for providing the MSI

images.

REFERENCES

Abdulla, A. and Ahmed, M. (2021). An improved image

quality algorithm for exemplar-based image inpaint-

ing. Multimed Tools Appl, 80:13143–13156.

Bertalmio, M., Sapiro, G., Caselles, V., and Ballester, C.

(2000). Image inpainting. In Proceedings of the 27th

Annual Conference on Computer Graphics and Inter-

active Techniques, SIGGRAPH ’00, page 417–424.

ACM Press/Addison-Wesley Publishing Co.

Easton, R. L., Christens-Barry, W. A., and Knox, K. T.

(2011). Spectral image processing and analysis of the

archimedes palimpsest. In 2011 19th European Signal

Processing Conference, pages 1440–1444.

Gippert, J. and Dum-Tragut, J. (2023). Caucasian Albania:

An International Handbook. De Gruyter.

Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and

Hochreiter, S. (2017). Gans trained by a two time-

scale update rule converge to a local nash equilibrium.

In Proceedings of the 31st International Conference

on Neural Information Processing Systems, NIPS’17,

page 6629–6640. Curran Associates Inc.

Jampour, M., Li, C., Yu, L.-F., Zhou, K., Lin, S., and

Bischof, H. (2017). Face inpainting based on high-

level facial attributes. Computer Vision and Image

Understanding, 161:29–41.

Jampour, M., Mohammed, H., and Gippert, J. (2023a).

Synthetic MSI Images of Georgian Palimpsests (SGP

Dataset).

Jampour, M., Zare, M., and Javidi, M. (2023b). Ad-

vanced multi-gans towards near to real image and

video colorization. J Ambient Intell Human Comput,

14:12857–12874.

Li, W., Lin, Z., Zhou, K., Qi, L., Wang, Y., and Jia, J.

(2022). Mat: Mask-aware transformer for large hole

image inpainting. In 2022 IEEE/CVF Conference on

Computer Vision and Pattern Recognition (CVPR),

pages 10748–10758. IEEE Computer Society.

Lucic, M., Kurach, K., Michalski, M., Bousquet, O., and

Gelly, S. (2018). Are gans created equal? a large-scale

study. In Proceedings of the 32nd International Con-

ference on Neural Information Processing Systems,

NIPS’18, page 698–707. Curran Associates Inc.

Obukhova, N. A., Baranov, P. S., Motyko, A. A.,

Chirkunova, A. A., and Pozdeev, A. A. (2023).

Palimpsest research based on hyperspectral informa-

tion. In 2023 25th International Conference on Digital

Signal Processing and its Applications (DSPA), pages

1–4.

c, T. and Pi

zurica, A. (2015). Context-aware patch-

based image inpainting using markov random ﬁeld

modeling. IEEE Transactions on Image Processing,

24(1):444–456.

Salerno, E., Tonazzini, A., and Bedini, L. (2007). Digi-

tal image analysis to enhance underwritten text in the

archimedes palimpsest. IJDAR, 9:79–87.

Seitzer, M. (2020). pytorch-ﬁd: FID Score for PyTorch.

https://github.com/mseitzer/pytorch- fid. Version

0.3.0.

Starynska, A., Messinger, D., and Kong, Y. (2021). Reveal-

ing a history: palimpsest text separation with genera-

tive networks. IJDAR, 24:181–195.

Xiang, H., Zou, Q., Nawaz, M. A., Huang, X., Zhang, F.,

and Yu, H. (2023). Deep learning for image inpaint-

ing: A survey. Pattern Recognition, 134:109046.

Zare, M., Lari, K. B., Jampour, M., and Shamsinejad, P.

(2019). Multi-gans and its application for pseudo-

coloring. In 2019 4th International Conference on

Pattern Recognition and Image Analysis (IPRIA),

pages 1–6.

Zhang, N., Ji, H., and et al., L. L. (2019). Exemplar-based

image inpainting using angle-aware patch matching. J

Image Video Proc, 70:1.

ICPRAM 2024 - 13th International Conference on Pattern Recognition Applications and Methods

694