SynFine: Boosting Image Segmentation Accuracy Through Synthetic

Data Generation and Surgical Fine-Tuning

Mehdi Mounsif

, Yassine Motie

, Mohamed Benabdelkrim

and Florent Brondolo

Akkodis Research, France

Keywords:

Computer Vision, Transfer Learning, Surgical Fine-Tuning, Synthetic Data Generation, Carbon Capture and

Storage.

Abstract:

Carbon Capture and Storage (CCS) has increasingly been suggested as one of the many ways to reduce CO

concentration in the atmosphere, hence tackling climate change and its consequences. As CCS involves robust

modelling of physico-chemical mechanisms in geological formations, it beneﬁts from CT-scans and accurate

segmentation of rock core samples. Nevertheless, identifying precisely the components of a rock formation can

prove challenging and could beneﬁt from modern segmentation approaches, such as U-Net. In this context, this

work introduces SynFine, a framework that relies on synthetic data generation and surgical ﬁne-tuning to boost

the performance of a model on a target data distribution with a limited number of examples. Speciﬁcally, after

a pre-training phase on a source dataset, the SynFine approach identiﬁes and ﬁne-tunes the most responsive

layers regarding the distribution shift. Our experiments show that, beyond an advantageous ﬁnal performance,

SynFine enables a strong reduction of the number of real-world labelled pairs for a given level of performance.

1 INTRODUCTION

In the recent years, many studies have highlighted the

strong correlation between human activity and global

warming, with CO

emissions being a major contrib-

utor to these dynamics (IPCC, 2022; Kramer et al.,

2021). As the highly likely environmental modiﬁ-

cations resulting from climate will have signiﬁcant

impacts on current societies and modern organisa-

tions, multiple institutions and actors have been de-

veloping policies, tools and methods (Kristj

ansd

ottir

and Kristj

ansd

ottir, 2021) to try and reduce the an-

thropogenic effects on the greenhouse gases (Ash-

worth et al., 2010; Huaman and Jun, 2014; Wenner-

sten et al., 2015) trapping the sun’s radiation within

the atmosphere. For instance, CCS (Carbon Capture

and Storage) technologies intend on capturing carbon

from emitters, and injecting its liquid form into deep

geological formations. While multiple geographic ar-

eas are already known for their signiﬁcant storage po-

tential, important efforts have been invested in esti-

mating the potential volume that can be cached in

porous rocks.

https://orcid.org/0000-0002-2763-3890

https://orcid.org/0000-0001-6191-6547

https://orcid.org/0000-0002-0978-3000

https://orcid.org/0000-0002-9168-8474

For this purpose, understanding the elaborate

physical, chemical and mechanical mechanisms in-

volved during CO

ﬂooding is of pivotal importance.

It is a complex task that requires accurate modelling

of the rock properties and is often based on the anal-

ysis and processing of CT-scanned volumes of rock

samples. In this context, image segmentation, a task

that consists in assigning a label to every pixel in an

image such that pixels belonging to a similar class

share certain characteristics, such as identifying all

the rock pores of a CT-volume, is a commonly used

approach to estimate rock porosity and adequately

guide the modelling phase. While crucial, segmen-

tation is, however, mostly done by hand and/or via

histogram-based readings, involving user bias and er-

rors, further leading to modelling inaccuracies (Saraf

and Bera, 2021).

In this view, the remarkable technical progresses

in computer vision, in particular through the increas-

ing usage of deep neural networks, has enabled the

scientiﬁc community to tackle tasks implying vari-

ous modalities and highly diverse environments. Nev-

ertheless, one of the drawbacks of these approaches

is the signiﬁcant amount of labelled data required to

train supervised models and ensure their robustness.

This is partially offset by techniques such as transfer

learning which, beyond offering a signiﬁcant training

Mounsif, M., Motie, Y., Benabdelkrim, M. and Brondolo, F.

SynFine: Boosting Image Segmentation Accuracy Through Synthetic Data Generation and Surgical Fine-Tuning.

DOI: 10.5220/0011848300003411

In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 565-573

ISBN: 978-989-758-626-2; ISSN: 2184-4313

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

565

time reduction, enables to reuse features learned from

an initial training task in other domains, making it a

highly relevant strategy in the scope of industrial us-

age, due to potentially complex data gathering and la-

belling processes.

In this context, this paper introduces SynFine, a

framework that enables the reduction of real-world la-

belled data needed for a given level of performance

through the usage of synthetic data and surgical ﬁne-

tuning, a branch of transfer learning that focuses on

speciﬁc parameters and challenges the common prac-

tice of training only the last layer. In particular, be-

yond shedding light on the strong synergy that can ex-

ist between transfer learning and synthetic data gen-

eration in an industrial scope, the contributions of this

work are the following:

• Validating the surgical ﬁne-tuning of a transfer

learning approach for segmentation tasks on a

dataset composed of corrupted images

• Providing a domain-driven example of procedural

data generation

• Experimentally evaluating the SynFine method

relevance regarding synthetic data generation and

empirically proving that signiﬁcant beneﬁts, in

particular labelling cost reduction, can be ob-

tained in this setting

2 RELATED WORKS

Accurate modelling of rock physics is a promising

and dynamic research direction (Ibrahim et al., 2021)

that could potentially lead to signiﬁcant workﬂow im-

provements in the geo-science industry and, in partic-

ular, CCS applications (Saraf and Bera, 2021). How-

ever, while notable progresses have been recently pre-

sented (Wang et al., 2018) in this particularly complex

ﬁeld, it is still typically approached using traditional

statistical methods and could arguably beneﬁt from

the use of contemporary deep learning ideas.

Surveying modern AI technical landscape, the re-

markable results and the highly diverse modalities

tackled by models are undeniable (Devlin et al., 2018;

Nichol et al., 2021) and hint at the possibilities of

these approaches to extend to a wide array of do-

mains, among which rock-physics modelling. One of

the main limitations of learning-based models is ac-

cess to qualitative and labelled data in sufﬁcient vol-

ume, which can be challenging and incur heavy costs

in practice (Everingham et al., 2015; Wu et al., 2021;

Christiano et al., 2017).

Given these constraints, transfer learning (Oquab

et al., 2014) has been one of the most popular strate-

gies employed to train more robust models and re-

duce overﬁtting on modestly sized dataset. While this

paradigm has been used through multiple domains

(Razavian et al., 2014), it relies essentially on an un-

changed workﬂow to adapt pre-trained features to tar-

get distributions. Speciﬁcally, multiple works (Kirk-

patrick et al., 2017; Lee et al., 2019) demonstrate ad-

vantageous performances when the classifying layers

of a model are ﬁne-tuned while the rest of the param-

eters are frozen (Sener et al., 2016; Kirichenko et al.,

2022). These strategies implement straightforward

mechanisms to prevent the loss of information and

have been extended with softer methods that involve

weight regularization and network pruning (Myung

et al., 2022). More recently, strong focus has been

directed towards few-shots adaptation (Shen et al.,

2021), general robustness (Andreassen et al., 2021)

and model adaptation (Varsavsky et al., 2020), such

as (Lee et al., 2022) on which this work heavily builds

and that introduces the idea of speciﬁc layer respon-

siveness given the shift between the initial and target

distribution.

The SynFine method presented in this work sug-

gests that this ﬁne-tuning paradigm can be particu-

larly relevant in the context of synthetic data gen-

eration (Nikolenko, 2021). Indeed, as opposed to

currently available manually-labelled datasets, which

may lack support for ﬁne-grained features and mostly

implement masks as rough patches (Lin et al., 2014)

due to the minutiae required, it is possible to generate

pixel-perfect image-mask pairs through simulation

and/or 3D modelling (Khan et al., 2019; Qiu et al.,

2017). This accuracy has encouraged many works to

rely on 3D pipelines as in (Johnson-Roberson et al.,

2016), where the authors show that using a relat-

able synthetic dataset several orders of magnitudes

more voluminous than its real counterpart enabled

their model to reach a higher score on the real-world

validation set. While this underlines the relevance of

pre-training on synthetic dataset, the work proposed

in this report suggests that potentially superior bene-

ﬁts can be reached through the introduction of speciﬁc

ﬁne-tuning strategies.

3 METHOD

As mentioned, the SynFine framework is heavily

geared towards industrial considerations and, as such,

suggests that the combined usage of surgical ﬁne-

tuning and synthetic datasets when real-world data is

scarce or costly to produce is a relevant solution. In

this view, this section ﬁrst presents the general Syn-

Fine pipeline and explains the stages required by this

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

566

method. Then, it provides detail regarding the data

generation process which may present some particu-

larities depending on the application context. Finally,

it introduces the RGN (Ratio Norm Gradient) metric

that can provide guidance regarding layer sensibility

to ﬁne-tuning and overall contribution to the model

performance in the downstream task.

3.1 The SynFine Pipeline

The SynFine framework is inspired by the increas-

ing interest in synthetic data generation approaches.

While the remarkable recent improvements in this

ﬁeld can not be understated, it is unlikely that sim-

ulations will be able to replicate exactly the complex

and highly non-linear processes and mechanisms un-

derlying the physical world.

However, if real-world data were to be consid-

ered as a corrupted version of data generated through

procedural functions, then it becomes possible to

frame this conﬁguration within the surgical ﬁne-

tuning paradigm. Speciﬁcally, following the ideas of

layer sensibilities to distribution shift in the scope of

transfer learning introduced in (Lee et al., 2022), Syn-

Fine, as presented in Figure 1 proposes the following

workﬂow:

1. Produce an abundant labelled synthetic dataset,

relatable to the target real-world dataset

2. Pre-train a given model on the synthetic data col-

lection

3. Identify the most suitable and sensitive layers re-

garding the distribution shift between the syn-

thetic dataset and the target real-world dataset

4. Apply surgical ﬁne-tuning to maximise the pre-

training beneﬁts and ﬁnal model performance in a

possibly limited data regime.

3.2 Data Generation

While the SynFine paradigm is completely generic,

there exist practical considerations that may signiﬁ-

cantly inﬂuence the data generation stage of the pro-

posed method. In particular, since this phase aims at

generating labelled synthetic data, it is crucial that the

processes setup in this regard provide a way to isolate

the mechanisms that produce the different labels re-

lated to the target dataset. As the current SynFine ap-

proach focuses mostly on expert-knowledge and man-

ually designed pipelines to create labelled variations

of images, as demonstrated in Section 4.1, more theo-

retical and causality-driven methods, for instance re-

lying on conditional generative modelling, will be ex-

plored in further work.

Figure 1: General view of the SynFine framework and the

main workﬂow steps.

3.3 Surgical Fine-Tuning and Layer

Sensibility Evaluation

As mentioned in Section 2, transfer learning has a

paramount importance in modern deep learning ap-

proaches and applications. While the traditional

workﬂow consists of an initial ﬁne-tuning of the

last layers and a progressive unfreezing of the ear-

lier blocks, the authors of (Lee et al., 2022) show

that, depending on the distribution shift between the

source and target dataset, this strategy can prove sub-

optimal and lead to lower performances than a surgi-

cal methodology that focuses the ﬁne-tuning on the

most suitable parameters.

Consequently, as suggested in Section 3.1, iden-

tifying layer sensibility is a crucial part of SynFine,

in particular since the alternative approach would re-

quire to evaluate the ﬁnal performance of each ﬁne-

tuning strategy, which is both inconvenient and in-

efﬁcient. In this view, different metrics can be con-

sidered to compute layer sensibility to the distribu-

tion shift between the synthetic dataset and the real-

world images. Speciﬁcally, the proposed pipeline has

been evaluated under the RGN responsiveness strat-

egy, which uses batches of data to estimate per-layer

contribution to the error relative to the layer weights.

Derived in Equation 1, with θ

and g

respectively be-

ing the parameters and the gradient of layer i.

RGN(θ

) =

||θ

(1)

SynFine: Boosting Image Segmentation Accuracy Through Synthetic Data Generation and Surgical Fine-Tuning

567

4 EXPERIMENTS

In order to demonstrate the applicability and rele-

vance of the SynFine framework, this section presents

the multiple experiments setup in this view as well

as the data created accordingly, central in this con-

text. Speciﬁcally, following details regarding the pro-

cesses used to generate mineral-inspired images and

their segmentation masks, layer sensibility metrics are

explored and the resulting performances for each ﬁne-

tuning strategies are analyzed to conﬁrm the interest

of the SynFine methodology.

For the experiments presented in this section, a

regular 3 blocks U-Net model is implemented in Py-

Torch, with 467k parameters. Training phases rely

on a Adam optimizer, with 2 × 10

−3

learning rate for

the initial pre-training phase and a 3 × 10

−4

learning

rate for the ﬁne-tuning. Images and masks are down-

sampled to 64×64 patches and a batch size of 32 is

used. Two thousand synthetic images were generated,

which is one order of magnitude more than the real-

world dataset. Finally, training and ﬁne-tuning phases

are both scheduled for 10 epochs.

4.1 Labelled Data Generation

While intrinsically generic, as mentioned in Section

3.2, some practical aspects of the target domain can

imply speciﬁc workﬂows regarding the data genera-

tion pipeline, as exempliﬁed in the following para-

graphs that detail the process of mineral-inspired la-

belled image generation.

Complex organic shapes observable in nature can

often be broken down into a non-linear combination

and superposition of simpler patterns. Accordingly,

the processing pipeline designed to implement these

principles, visible in Figure 2, involves tiling multiple

basic shapes while applying per-shape stretches and

random scaling, rotations and offsets. These transfor-

mations are applied to each basic unit and the result-

ing collections of patterns are then blended using a

max-intensity strategy which, contrary to an additive

strategy, prevents saturating the pixels as this seems

to be a notable feature of the target class.

While this ﬁrst function provides the image back-

ground, the target class pixels can be created by ap-

plying a threshold function to the histogram of a noise

texture as a starting image, consequently generating

large areas without information and specks of clearer

pixels. The application of a gaussian bluring ﬁlters

can contribute in diffusing these pixels in wider areas

which can then go through a step function to provide

the desired masks. As feeding seeds to the noise gen-

erator will result in different masks, this approach is

a convenient way to produce a large dataset for rock

segmentation.

4.2 Surgical Fine-Tuning Validation on

Corrupted Dataset

The main ideas of surgical ﬁne-tuning were initially

introduced in the scope of classiﬁcation tasks. In

these cases, input-level perturbations, among other

types of distribution shift, were shown to strongly af-

fect native performance. In this context, the ﬁrst step

towards the validation of the SynFine framework was

to evaluate the system behaviour in a segmentation

task with similar input corruptions.

As such, in this early-stage validation phase, no

synthetic data is required since it rather aims at eval-

uating surgical ﬁne-tuning technique in a segmenta-

tion context. In practice, as shown in Figure 3, cor-

rupted data is derived from the initial rock segmenta-

tion dataset through the application of various visual

perturbations.

After the initial training phase on the real dataset,

the RGN sensibility metric, introduced in Section 3.3,

is computed in order to provide guidance to the surgi-

cal ﬁne-tuning phase. In practice, two ways of clus-

tering the model parameters are analyzed due to the

U-Net architecture speciﬁcity:

• The sequential view where blocks are considered

independently, that is, all the encoding blocks fol-

lowed by the decoding ones

• The transversal view, relying on a horizontal clus-

tering of blocks the encoding/decoding pairs. In

this case, parameters are gathered depending on

their depth in the model.

Figure 4 displays the sensibility of each parameter

group for both clustering paradigms. As can be seen,

in both cases, the ﬁrst/higher U-Net block indicates

the highest responsiveness, which is coherent with the

expected results since the perturbations added focus

on input-level features. Since multiple batches of data

are used, the shaded area represent the min-max vari-

ations while the solid line indicates the mean layer

responsiveness value.

Finally, the ﬁnal ﬁne-tuning results for different

freezing strategies, visible in Figure 5, conﬁrm that

surgical ﬁne-tuning:

• Signiﬁcantly outperforms naive ﬁne-tuning

• Strongly inﬂuences the ﬁnal model performance

based on the parameters added to the trainable set

• Is also relevant in a segmentation context, despite

the introduction of transversal gradients paths, in-

herent to the U-Net architecture.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

568

Figure 2: Mineal-inspired image generation process and reference image.

Figure 3: Instance of a real-world input image and multiple corruptions examples.

Figure 4: Layer responsiveness after initial training for

RGN metrics on a corrupted version of the real-world

dataset.

4.3 Transfer from Procedural Data to

Real-World Images

This second experiment focuses on the main use case

for the SynFine framework. Speciﬁcally, it considers

the case where an abundant labelled synthetic dataset

is available on which a segmentation model is ini-

tially trained. Then, using layer sensibility evaluation,

the most responsive parameters regarding the distri-

bution shift are identiﬁed and ﬁne-tuned on the real-

world rock segmentation dataset. In this view, Figure

6 shows the resulting accuracy levels for various ﬁne-

tuning strategies, along with the RGN sensibility per

layer for both sequential and transversal view. While

a range of ﬁnal performances can be observed, every

surgical ﬁne-tuning strategy strongly outperforms the

native model accuracy (that is, without ﬁne-tuning) as

well as the naive approach that consists in ﬁne-tuning

only the last layer, as summed up in Table 1.

It is however less trivial to interpret the RGN sen-

sitivity measure, in particular in the Sequential con-

ﬁguration. Speciﬁcally, while the highest RGN scor-

ing layer does indeed signal the most responsive pa-

rameters in both cases and follows the general shape

of the surgical ﬁne-tuning performance, the relation

between RGN values and ﬁnal scores of the Down-

Block 1, 2 and UpBlock 0 is not entirely clear and

will be further investigated in subsequent works.

Table 1: Surgical ﬁne-tuning approaches performance in-

crease with native model and naive ﬁne-tuning.

Accuracy

relative to

DownBlock 0 Trans. High

Native performance 26.6 ± 0.20 28.5 ± 0.27

Naive ﬁne-tuning 35.9 ± 0.20 37.8 ± 0.28

SynFine: Boosting Image Segmentation Accuracy Through Synthetic Data Generation and Surgical Fine-Tuning

569

Figure 5: Accuracy comparison for different surgical ﬁne-tuning target for both the sequential and horizontal clustering. From

left to right, DownBlock 0-3, UpBlock 0-1 and the three depth levels of the transversal view.

Figure 6: RGN layer responsiveness in a synthetic-to-real transfer for both parameters clustering paradigms.

4.4 Synthetic Advantages

The previous sections have provided experimental ev-

idence of the relevance of surgical ﬁne-tuning strate-

gies for segmentation tasks and conﬁrmed that cor-

rectly selecting which model parameters should be

frozen during the optimization phase depending on

the distribution shift can yield non-negligible increase

in performance, contrary to naive approaches.

It is, however, also interesting to view the SynFine

framework in the context of dataset creation. Indeed,

experiments presented in Sections 4.2 and 4.3 focus

on ﬁnal performance for a given dataset size and vol-

ume. Nevertheless, in the rock segmentation conﬁgu-

ration for instance, an operating company could very

likely be aiming at minimizing the number of samples

required to reach a given level of performance. In this

paradigm, the per-sample efﬁciency is paramount and

the SynFine approach can provide signiﬁcant advan-

tage, as can be seen in Figure 7. Speciﬁcally, three

approaches are compared:

• A straightforward strategy, named Baseline and

shown in violet, which does not rely on any form

of transfer learning

• A naive transfer learning approach that ﬁne-tunes

only the last layer, labelled Traditional FT and dis-

played in blue

• Surgical FT, the responsiveness-driven surgical

ﬁne-tuning approach, in red.

As can be observed, the SynFine method displays

a higher initial accuracy on the target dataset and

shows that, in this conﬁguration, labelled samples

yield a more advantageous increase in performance

than with other methods, thus underlining the strong

interest of this approach in domains where data can be

simulated but labelling is expensive. While the base-

line performance gets close to the surgical ﬁne-tuning

when approaching the full dataset, it is likely that the

gap could be more important if the images considered

were more complex and will be explored in further

works.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

570

Figure 7: Segmentation accuracy comparison for multiple ﬁne-tuning strategies against a baseline process for increasing

dataset size (x-axis).

5 CONCLUSIONS

In this work, the SynFine framework, an empirical

approach for maximizing transfer learning beneﬁts

for segmentation tasks in the scope of synthetic data

generation is introduced. Through multiple experi-

ments, displaying synthetic data generation strategies

for real-world data distributions in different contexts,

it has been shown that there exist signiﬁcant advan-

tages regarding performance when leveraging surgi-

cal ﬁne-tuning. Additionally, this work provides local

evidence that the SynFine framework can prove con-

siderably more cost-effective than a naive approach,

which is a central aspect in many real-world indus-

tries and applications. In particular, the improvements

induced by the SynFine method could impulse sig-

niﬁcant advantages regarding CCS modelling and, in

ﬁne, deployment.

While having direct practical implications, the ini-

tial results presented in this work also open a vast ar-

ray of perspectives. Among other considerations, it

would be insightful to understand how to design syn-

thetic data that triggers essentially intermediate U-Net

blocks or levels, since most of the sensibility in our

experiments was concentrated within the ﬁrst block.

This understanding could further be leveraged to help

formulate a more theoretical compatibility measure

between the target real-world dataset and the synthet-

ically produced, which is lacking from the current ap-

proach.

Finally, regarding the synthetic data generation,

while the presented strategy relies on expert knowl-

edge to mass-produce labelled data, the increasing ca-

pabilities of generative models, in particular within

the diffusion scope, and their ability to produce

conditionally-driven images could also be considered

to complement or completely replace the procedural

approaches presented in this work and also provide

more accessibility for domain of higher complexity.

ACKNOWLEDGEMENTS

This work has been sponsored by the Akkodis group.

REFERENCES

Andreassen, A., Bahri, Y., Neyshabur, B., and Roelofs, R.

(2021). The evolution of out-of-distribution robust-

ness throughout ﬁne-tuning. ArXiv, abs/2106.15831.

Ashworth, P., Boughen, N., Mayhew, M., and Millar, F.

(2010). From research to action: Now we have to

move on ccs communication. International Journal

of Greenhouse Gas Control, 4(2):426–433.

Christiano, P. F., Leike, J., Brown, T., Martic, M., Legg, S.,

and Amodei, D. (2017). Deep reinforcement learning

from human preferences. In Guyon, I., Luxburg, U. V.,

Bengio, S., Wallach, H., Fergus, R., Vishwanathan,

S., and Garnett, R., editors, Advances in Neural Infor-

SynFine: Boosting Image Segmentation Accuracy Through Synthetic Data Generation and Surgical Fine-Tuning

571

mation Processing Systems, volume 30. Curran Asso-

ciates, Inc.

Devlin, J., Chang, M.-W., Lee, K., and Toutanova,

K. (2018). Bert: Pre-training of deep bidirec-

tional transformers for language understanding. cite

arxiv:1810.04805Comment: 13 pages.

Everingham, M., Eslami, S. M. A., Van Gool, L., Williams,

C. K. I., Winn, J., and Zisserman, A. (2015). The pas-

cal visual object classes challenge: A retrospective.

International Journal of Computer Vision, 111(1):98–

136.

Huaman, R. N. E. and Jun, T. X. (2014). Energy related co2

emissions and the progress on ccs projects: a review.

Renewable and Sustainable Energy Reviews, 31:368–

385.

Ibrahim, E., Jouini, M., Bouchaala, F., and Gomes, J.

(2021). Simulation and validation of porosity and

permeability of synthetic and real rock models using

three-dimensional printing and digital rock physics.

ACS Omega.

IPCC (2022). Summary for Policymakers. Cambridge Uni-

versity Press, Cambridge, UK. In Press.

Johnson-Roberson, M., Barto, C., Mehta, R., Sridhar, S. N.,

and Vasudevan, R. (2016). Driving in the matrix: Can

virtual worlds replace human-generated annotations

for real world tasks? CoRR, abs/1610.01983.

Khan, S., Phan, B., Salay, R., and Czarnecki, K. (2019).

Procsy: Procedural synthetic dataset generation to-

wards inﬂuence factor studies of semantic segmenta-

tion networks. In Proceedings of the IEEE/CVF Con-

ference on Computer Vision and Pattern Recognition

(CVPR) Workshops.

Kirichenko, P., Izmailov, P., and Wilson, A. G. (2022). Last

layer re-training is sufﬁcient for robustness to spurious

correlations.

Kirkpatrick, J., Pascanu, R., Rabinowitz, N., Veness, J.,

Desjardins, G., Rusu, A. A., Milan, K., Quan, J.,

Ramalho, T., Grabska-Barwinska, A., Hassabis, D.,

Clopath, C., Kumaran, D., and Hadsell, R. (2017).

Overcoming catastrophic forgetting in neural net-

works. Proceedings of the National Academy of Sci-

ences, 114(13):3521–3526.

Kramer, R. J., He, H., Soden, B. J., Oreopoulos,

L., Myhre, G., Forster, P. M., and Smith, C. J.

(2021). Observational evidence of increasing global

radiative forcing. Geophysical Research Letters,

48(7):e2020GL091585.

Kristj

ansd

ottir, H. and Kristj

ansd

ottir, S. (2021). Carbﬁx

and sulﬁx in geothermal production, and the blue la-

goon in iceland: Grindav

ık urban settlement, and vol-

canic activity. Baltic Journal of Economic Studies,

7(1):1–9.

Lee, J., Tang, R., and Lin, J. (2019). What would elsa do?

freezing layers during transformer ﬁne-tuning. ArXiv,

abs/1911.03090.

Lee, Y., Chen, A. S., Tajwar, F., Kumar, A., Yao, H., Liang,

P., and Finn, C. (2022). Surgical ﬁne-tuning improves

adaptation to distribution shifts.

Lin, T., Maire, M., Belongie, S. J., Bourdev, L. D., Girshick,

R. B., Hays, J., Perona, P., Ramanan, D., Doll

ar, P.,

and Zitnick, C. L. (2014). Microsoft COCO: common

objects in context. CoRR, abs/1405.0312.

Myung, S., Huh, I., Jang, W., Choe, J. M., Ryu, J., Kim,

D., Kim, K.-E., and Jeong, C. (2022). Pac-net: A

model pruning approach to inductive transfer learning.

In International Conference on Machine Learning.

Nichol, A., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin,

P., McGrew, B., Sutskever, I., and Chen, M. (2021).

GLIDE: towards photorealistic image generation and

editing with text-guided diffusion models. CoRR,

abs/2112.10741.

Nikolenko, S. (2021). Synthetic Data for Deep Learning.

Springer Optimization and Its Applications. Springer

International Publishing.

Oquab, M., Bottou, L., Laptev, I., and Sivic, J. (2014).

Learning and transferring mid-level image represen-

tations using convolutional neural networks. In 2014

IEEE Conference on Computer Vision and Pattern

Recognition, pages 1717–1724.

Qiu, W., Zhong, F., Zhang, Y., Qiao, S., Xiao, Z., Kim,

T. S., and Wang, Y. (2017). UnrealCV: Virtual worlds

for computer vision. In Proceedings of the 25th

ACM International Conference on Multimedia, page

1221–1224. Association for Computing Machinery.

Razavian, A. S., Azizpour, H., Sullivan, J., and Carlsson,

S. (2014). Cnn features off-the-shelf: an astounding

baseline for recognition.

Saraf, S. and Bera, A. (2021). A review on pore-scale mod-

eling and ct scan technique to characterize the trapped

carbon dioxide in impermeable reservoir rocks during

sequestration. Renewable and Sustainable Energy Re-

views, 144:110986.

Sener, O., Song, H. O., Saxena, A., and Savarese, S.

(2016). Learning transferrable representations for un-

supervised domain adaptation. In Lee, D., Sugiyama,

M., Luxburg, U., Guyon, I., and Garnett, R., editors,

Advances in Neural Information Processing Systems,

volume 29. Curran Associates, Inc.

Shen, Z., Liu, Z., Qin, J., Savvides, M., and Cheng,

K.-T. (2021). Partial is better than all: Revisiting

ﬁne-tuning strategy for few-shot learning. ArXiv,

abs/2102.03983.

Varsavsky, T., Orbes-Arteaga, M., Sudre, C. H., Graham,

M. S., Nachev, P., and Cardoso, M. J. (2020). Test-

time unsupervised domain adaptation. In Martel,

A. L., Abolmaesumi, P., Stoyanov, D., Mateus, D.,

Zuluaga, M. A., Zhou, S. K., Racoceanu, D., and

Joskowicz, L., editors, Medical Image Computing

and Computer Assisted Intervention. Springer Inter-

national Publishing.

Wang, B., Chen, Y., Lu, J., and Jin, W. (2018). A rock

physics modelling algorithm for simulating the elastic

parameters of shale using well logging data. Scientiﬁc

Reports, 8.

Wennersten, R., Sun, Q., and Li, H. (2015). The future

potential for carbon capture and storage in climate

change mitigation–an overview from perspectives of

technology, economy and risk. Journal of Cleaner

Production, 103:724–736.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

572

Wu, J., Ouyang, L., Ziegler, D. M., Stiennon, N., Lowe, R.,

Leike, J., and Christiano, P. F. (2021). Recursively

summarizing books with human feedback. ArXiv,

abs/2109.10862.

SynFine: Boosting Image Segmentation Accuracy Through Synthetic Data Generation and Surgical Fine-Tuning

573