ITM Web of Conferences 64, 01010 (2024) https://doi.org/10.

1051/itmconf/20246401010
ICACS24

2D to 3D Image Conversion Algorithms

Samar Sabah Mamand1*, Abdulqadir Ismail Abdulla2
1Computer Science Department, College of Science, Salahaddin University, Kurdistan Region,
Erbil, Iraq
2Computer Engineering Department, College of Engineering, Knowledge University, Kurdistan

Region, Erbil, Iraq.

Abstract. With the emergence of Artificial Intelligence (AI), there are

many applications for 3D computer vision, and different problems in
diversified domains are being solved. Particularly deep learning and image
processing techniques are widely used in computer vision applications, for
example, medical imaging which commonly uses 2D images to see human
organs can benefit tremendously from 3D reconstruction of a human organ
or cancer lesions for diagnosis purposes. Cars nowadays use radar, lidar, and
sensors to create an awareness of the surroundings of the car, having a 3D
reconstruction of the car's surroundings will provide a better situational
awareness. In many computer vision applications, it is essential to use image
conversion techniques. With AI in place, learning-based approaches became
popular along with image processing techniques. Towards this end, this
paper throws light on those methods with a systematic literature review. The
insights presented in this paper can help in further research in computer
vision applications.

1 Introduction
Nowadays, one of the core challenges of image-based modeling and computer vision is
producing a realistic 3D model from 2D photos. There are numerous algorithms available
today for reconstructing 3D from 2D photos, techniques use only a single image as input
(Binocular disparity, Motion Parallax, Image blur, Silhouette, and Structure from motion),
and techniques use Two or more than two images (Linear perspective, Atmosphere scattering
and Shape from shading) each technique has unique execution requirements, advantages, and
disadvantages. Particularly image content is widely used in different applications of
computer vision and image processing. Images are widely used in real-time applications. For
instance, applications about human identity use photos (faces) to identify humans. In some
applications, images are used to preserve historical activities or happenings for future
revision and retrieval. Digital image processing is the approach in which digital computers
are used to deal with image processing. In image processing, there is some process involved
based on the problem at hand. For each problem associated with image processing specific
algorithm is used.
For instance, image segmentation is done by using segmentation-based algorithms. Image
processing therefore is based on several algorithms. Images can be in either 2D or 3D form.
In many real-life applications 2D and 3D images are being used. In the process of using them,
sometimes, it is important to convert from 2D to 3D and vice versa [1].

*
Corresponding author: [email protected]

© The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons
Attribution License 4.0 (https://creativecommons.org/licenses/by/4.0/).
ITM Web of Conferences 64, 01010 (2024) https://doi.org/10.1051/itmconf/20246401010
ICACS24

There are many techniques available for image conversion. Traditional image processing
approaches are used to convert from 2D to 3D. These approaches are based on image
processing algorithms. The conversion process involves certain heuristics and methods.
Another category of conversion techniques is learning-based approaches. Learning-based
approaches such as RANSAC (RAndom SAmple Consensus) and (3D Lider) are very
dynamic and they learn from available data rather than following certain common heuristics
[13].

Therefore, it is important to understand different conversion techniques. This paper's goal is

to examine the state of the art in terms of methods for converting 2D to 3D images. The main
objective is to review existing methods, their advantages, and limitations. Our contributions
in this work include a thorough analysis of the literature to determine the state of the art for
image conversion methods at the moment. This review provides useful insights about image
conversion methods. The remainder of the paper is structured as follows. Section 2 presents
our methodology used for a systematic review of state of the art on different techniques used
for 2D to 3D conversion and vice versa. Section 3 reviews the literature on 2D to 3D
conversion techniques. Section 4 provides a summary of findings and comparisons. Section
5 provides conclusions drawn from the review.

2 Methodology
Artificial Intelligence has been used widely in the medical field. It ranges from medical
diagnosis to treatment. It even helped sometimes in replacing the traditional medical
procedures because it avoids the basic errors that can be made by humans. The following
sections are the summery offered by this research in the major medical areas that can benefit
from the application of artificial intelligence.
Table 1: Shows final list of article references and their publisher

Publisher References
IEEE 4
Elsevier 9
Springer 2
Google Scholar 1

3 Literature Review
This section reviews the literature on different methods associated with 2D to 3D image
conversion. It provides insights on various aspects of conversion methods including machine
learning and deep learning.

3.1 Image Conversion Using Deep Learning

Deep learning models are widely used for processing image data. Addressed challenges in
producing a 3D form of porous media from 2D form, emphasizing the limitations of existing
methods. It introduces a novel deep learning-based framework for efficient and accurate 3D
reconstruction, achieving a remarkable speedup compared to traditional methods. The
proposed approach learns 2D to 3D form using neural networks, demonstrating instant
reconstruction with only 0.2s required for a 1283 reconstruction. The framework allows

2
ITM Web of Conferences 64, 01010 (2024) https://doi.org/10.1051/itmconf/20246401010
ICACS24

diversity to enhance adaptability. Experimental evaluations on isotropic and non-stationary

materials show the method's accuracy, speed, and stability. Despite visual imperfections and
challenges in training, the paper discusses potential improvements, including incorporating
user-defined constraints, and classical features, and addressing memory constraints for larger
image sizes. The study highlights the improvement of integrating the classical and neural
network-based features [6].

3.1.1 Transfer Learning

The eight 3D microstructures that are reconstructed using the provided method fall into many
microstructure categories (such as porous, composite, etc.) and each has distinct
characteristics that the methodology can accurately identify and rebuild. Based on the type
of image, such as RGB, binary, or grayscale, the sample sets are arranged. It is obtained
VGG19 (Visual Geometry Group — 19) is an ImageNet-trained deep convolutional neural
network (CNN), It was used to build a 3D microstructure. Only 2D photos with RGB
channels can be analyzed by VGG19. We add a permutation operator to VGG19's structure
before the first convolution layer to enable the use of 3D images with it, correlated 2D
characteristics are produced by VGG19. Deterministic features, such as some 2D response
maps, are generated by VGG19. We compute the Gram matrices associated with these
response maps and use them in the objective function formulation to translate these features
into statistical measures and guarantee solution existence during optimization [2].

3.1.2 Deep Fake Construction

The approach used with pictures and videos is called a deep fake. Offered a new method that
uses deep fakes to determine created movies and photos. Deep fake construction is a novel
way to reveal artificial intelligence-generated phony face pictures or videos, or "Deep Fakes."
The discovery that generated face regions are spliced into original images to create Deep
Fakes is the basis for our approach. With the original person's facial expression intact, these
algorithms generate new faces for them. Yet, because the neural network synthesis technique
does not ensure that the original face and the synthetic face have consistent facial landmarks,
the two faces have mismatched facial landmarks, which are positions on human faces
correlating to crucial structures like eye and mouth points. While the inaccuracies in
landmark placements might not be immediately apparent to the human eye, they can be
discerned using head poses, or the orientation and position of the head, which are calculated
from 2D landmarks in both the real and simulated facial regions. To train a classifier and
efficiently discriminate between original and Deep Fake photos, variations in generated head
positions are recorded and used as a vector of characteristics. The algorithm's effectiveness
in identifying AI-generated false face content is demonstrated via experimental evaluations
[16].

3.1.3 Voxel-Based Three-View Approach

Voxel representation of image data became important in computer vision applications.
Introduced a hybrid methodology for the classification of 3D shapes towards making a 3D-
based technology. Traditionally, methods converted 2D convolution into 3D form for binary
voxel representation, potentially losing crucial internal information for recognition. The
proposed approach obtains depth projection views from other poses and preserves
information about spatial. The defined network, combined with a voxel sub-network,
undergoes a fusion of weights, followed by classification. Experimental results on
ModelNet10 and ModelNet40 demonstrate competitive performance. [4]. In a world
increasingly reliant on three-dimensional models, the paper highlights the significance of

3
ITM Web of Conferences 64, 01010 (2024) https://doi.org/10.1051/itmconf/20246401010
ICACS24

accurate classification and retrieval, emphasizing the advantages and challenges posed by the
three-dimensional spatial structure. The proposed algorithm showcases effectiveness, and
promising applications in various fields, and future research will focus on computational
complexity and cost considerations to enhance deployment ability [4].

3.1.4 CNN-Based Methodology

Convolutional Neural Network (CNN) is found appropriate to process images. introduced a
novel 3D cell segmentation approach, addressing challenges in error-free automatic
segmentation of cellular membranes linked to 3D microscopic image data with variations in
intensities. The proposed method combines the power of CNN for pre-processing and
explores the efficacy of novel strategies for post-processing. By interpreting the segmentation
issues from a multi-instance perspective to a semantic perspective, the algorithm leverages
CNNs for optimal watershed algorithm application. With the usage of 3D images that are
annotated, the proposed method demonstrates better performance compared to existing
approaches. The pipeline enhances quality in segmentation in deep layers, eliminates tedious
parameter adjustments, produces dense masks for separation of foreground from background,
and helps in identifying the seed points involved in the method. Future improvements aim to
refine probability maps through additional training data for closer alignment to achieve error-
free automatic 3D segmentation [5].

3.1.5 Using Adversarial Networks

Adversarial networks are used to leverage training quality and improve model performance,
delved into the realm of Generative Adversarial Networks (GANs) in the context of image
synthesis. GANs have seen significant success across different domains.
The paper addresses the lack of a comprehensive review by covering GANs' loss variants
training and image generation. It categorizes synthetic image generation methods, exploring
image translation, image fusion, mapping of images, and converting text to image. The
methodology is based on learning models, ideas with architecture, datasets, and evaluation
approaches. The review highlights milestones, presents insights on development routes, and
outlines potential future research directions. Notably, different GAN approaches were
compiled and made accessible, providing a valuable resource for further exploration. The
paper also emphasizes the importance of explicit loss definition for image synthesis tasks,
addressing the challenges and strengths of various supervised and unsupervised methods. The
review concludes by pointing out the need for more research on domain adaptation, transfer
learning, and the impact of the utility of CNNs [11].

3.1.6 Micro Deep Learning Profilometry Approach

Micro deep learning is the attempt to improve the learning process, the quest for obtaining
rich object information rapidly and accurately from optical signals has driven advancements
in imaging technologies. Despite ultra-fast photography reaching speeds exceeding one
quadrillion frames per second, it remains limited to recording two-dimensional images,
lacking depth information crucial for understanding complex real-world objects. Addressing
this limitation, the paper introduces 𝜇DLP, a high-speed three-dimensional (3D) surface
imaging approach utilizing structured light illumination. Through a learning deep model,
𝜇DLP predicts desired information from the image, enabling conversion into 3D shapes.
Experiments demonstrate 𝜇DLP's capability to faithfully retrieve dynamic object geometry
at 20,000 frames per second, surpassing 3D approaches with acceptable accuracy. The
method contributes to bridging the gap between 3D and 2D imaging, offering advantages for
diverse use cases [7].

4
ITM Web of Conferences 64, 01010 (2024) https://doi.org/10.1051/itmconf/20246401010
ICACS24

3.2 Converting to 3D Super-Resolution Images

Image conversion plays a crucial role in computer vision applications. focused on super-
resolution imaging techniques that have enhanced spatial resolution in light microscopy but
often face constraints like specialized sample preparation. Obtaining high-speed, 3D
information within live cells remains a challenge. The SPEED (single-point edge-excitation
sub-diffraction) microscopy, a 2D single-molecule tracking technique, overcomes limitations
and provides a feasible solution. Leveraging a 2D-to-3D conversion method, SPEED
converts fast-captured 2D localization data into 3D super-resolution information. This
method, applicable to standard laboratory microscopes, allows cost-effective and
scientifically viable acquisition of super-resolution 3D data. Existing challenges in imaging
3D structures are addressed through optical sectioning and post-imaging transformations,
with a brief review of available techniques. SPEED's 2D-to-3D transformation offers an
accessible approach for deriving 3D forms from 2D, showcasing its potential for studying
various cellular structures. Future refinements aim to apply SPEED to asymmetric structures
like endosomes and mitochondri [8].

3.3 Regional Feature Fusion Method

Feature fusion is used in image processing to improve performance in problem-solving.
Autonomous vehicles demand precise perception systems for informed decision-making.
This study introduces a novel approach merging 2D and 3D object detection to enhance
accuracy in autonomous vehicles. It projects 3D-LiDAR data into image space, utilizes RPN
for ROI, and employs ROI pooling and Faster-RCNN for object detection. Fusing 3D-LiDAR
and camera features improves dimension prediction, overcoming scale changes and
occlusions. Assessment of the KITTI dataset demonstrates superior performance (94.59%
car, 82.50% van, 79.60% truck, 85.31% pedestrian, 86.33% cyclist) and real-time processing
(70 ms). Autonomous vehicles, relying on sensors like LiDAR and cameras, benefit from this
method for reliable tracking and decision-making, essential for safety and comfort
improvements [13].

3.4 3D Face Reconstruction Method

Face reconstruction is used in some computer vision applications dealing with identity.
Explored the 3D face generation which is vital for multimedia applications, often relying on
CNN-based models. However, limited annotated 3D training data poses challenges. To
address this, the proposed 2D-Assisted Learning (2DAL) leverages "in the wild" 2D with
landmark features, significantly enhancing 3D face model learning. Introducing four self-
supervision schemes based on landmark self-prediction consistency and cycle consistency,
2DAL effectively utilizes sparse 2D facial landmark heatmaps. These schemes reduce the
need for paired annotations for conversions, resulting in higher-quality 3D face models.
Experimentation on different datasets demonstrates that 2DAL outperforms existing methods
in both 3D face production and alignment of the face, offering promising advancements in
computer vision and multimedia application [12].

3.5 Voxel Based Representation Approach

Voxel representation of images is useful in computer vision applications, the utilization of
point cloud using vision techniques, has become prevalent in various applications,
particularly for mapping 3D urban scenes. Despite their efficacy in providing spatial
coordinates of geometric surfaces, raw point clouds lack structured information, hindering
their pre-processing and application. To address this, organizing 3D discrete points into
voxels, analogous to pixels in images, is a common strategy. This paper conducts a

5
ITM Web of Conferences 64, 01010 (2024) https://doi.org/10.1051/itmconf/20246401010
ICACS24

comprehensive review of representations of point clouds based on the voxel-based method,

emphasizing creation, utilization, and strengths/weaknesses. It explores their potential in the
construction industry, discussing applications like monitoring construction sites, navigation,
and as-built model reconstruction. The review highlights voxel-based representations'
advantages in data compression, storage, indexing, and 3D space description, offering a
powerful tool for diverse applications despite some limitations [15].

3.6 Advanced 3D Motion Prediction

Compression of images is widely used in many image processing applications, addressing
challenges in point cloud compression for immersive media representation, particularly
focusing on the MPEG-I standard. The current methods, such as 3D-based and 2D-video-
based approaches, encounter issues like varying point counts and lack of explicit
correspondence, hindering efficient motion estimation. The proposed model introduces a 3D
to 2D motion model, leveraging 3D motion to derive accurate 2D motion vectors (MVs).
Additionally, novel methods are introduced for precise 2D MV estimation, addressing issues
in patch inconsistency. The solutions are implemented demonstrating significant coding
gains compared to existing methods. The proposed algorithm effectively utilizes geometry
and auxiliary information, showcasing quality improvements in compression under the
MPEG-I framework. Future research will explore enhanced video compression algorithms
for V-PCC [9].

3.7 Mesoscopic Modelling

Microscopic modeling in image processing has its significance in inaccurate representation,
exploring the mesoscale modeling of concrete, focusing on its mechanical responses under
various loadings. Given the variety of concrete, it emphasizes the 3D random mesoscale
modeling approach. The study uses the 3D approach to examine the static and dynamic
behaviors of regular concrete under compression, split, and direct tension. The suggested
method shows great validity in describing the mechanical reactions of concrete, providing
insightful information for the creation and use of concrete buildings and materials in a variety
of contexts. In addition, the paper addresses the significance of examining concrete's both
static and dynamic behaviors given its extensive use in construction; it offers a thorough
analysis of mesoscale modeling techniques and compares their efficacy; it emphasizes the
need to use nonlinear material models to accurately describe concrete components; and it
concludes that the developed 3D random mesoscale modeling approach is dependable for
visualizing concrete's behaviors, enhancing understanding and application in engineering
[14].

3.8 The 3D Modular Chaotic Map

Chaotic map theory is used in image processing to improve the solution of a problem. used a
powerful image encryption in three dimensions, highlighting various modes of data integrity.
The suggested approach employs multiple techniques to enhance image quality. The color
image is converted to 3D space by splitting it up into smaller images according to RGB
spectra. The encryption process consists of permutation operations with a 3D modular chaotic
map, 3D operations, XOR, and circular shift operators. The 3D modular chaotic map that is
shown improves key space, speed, and period. Both the resilience and accuracy metrics are
enhanced by the method. The paper highlights the importance of cryptography in the context
of information security within computerized networks for using numerous qualities. The
proposed color image encryption approach demonstrates competitive security performance
with high efficiency, supporting various analyses such as keyspace, histogram, and

6
ITM Web of Conferences 64, 01010 (2024) https://doi.org/10.1051/itmconf/20246401010
ICACS24

correlation. The method's applicability extends to square and scalable color images, offering
potential use in key generation, watermarking, and steganography techniques [3].

3.9 Photo-Sketching Method

The photo-sketching method is found useful in law-enforcing applications, focused on
generating contour drawings, emphasizing their significance in computer graphics and
computer vision. Contour drawings, capturing object outlines and salient edges, convey 3D
perspective and are essential for understanding scene geometry. Previous approaches treated
this problem as boundary detection, but discrepancies in visual cues and the lack of artistic
style were observed. The paper introduces a dataset that facilitates diverse annotations and
imperfect alignment with ground truth. The study explores contour drawings as an
intermediate representation between image boundaries and abstract line drawings,
contributing to machine understanding and the generation of abstract visuals [10].

4 Discussion and Comparison

This section analyses a summary of the selected studies that have been reviewed in Section
3. The existing methods found for 2D to 3D conversion are summarized in Table 2, and Table
3 in terms of deep learning and image processing techniques used.

Table 2: Summary of existing methods used for 2D to 3D image conversion

Reference Approach Technique Algorithm Data set Limitation / Future

Scope
[1] Deep learning GAN Optimization- Custom Compression and
and layer-by- based and multi- dataset pruning-based
layer point statistics- improvements are still
reconstruction based algorithms desired.
approaches

[2] Deep learning CNN Lattice-point ImageNet In the future,

backpropagation reconstruction costs are
algorithm yet to be decreased.

[4] Deep learning CNN and RANSAC KITTI A more reliable and
sliding object accurate approach is yet
window detection desired.
technique benchmark
dataset
[6] Deep learning Laser Flow- Custom Voxel and point cloud-
scanning and constrained dataset based approaches are to
stereo vision clustering be investigated in the
techniques algorithm future.

[10] Deep learning, 3D object SVM ModelNet10 Computational cost and

GVCNN, and classification and computational
BPS-Conv3D ModelNet40 complexity are to be
reduced in the future.

[11] Deep learning CNN Seeded Custom The accuracy of the

watershed-based dataset model is to be improved
algorithms with additional training
data.

7
ITM Web of Conferences 64, 01010 (2024) https://doi.org/10.1051/itmconf/20246401010
ICACS24

[14] Deep Learning 3D Phase Custom The proposed method

surface demodulation dataset needs to be improved
imaging algorithm with the introduction of
efficient signal
processing.
[15] Deep Learning Salient Contour BSDS500 A more scalable
boundary generation alternative is yet desired.
detection algorithm
[7] Machine DeepFake Levenberg- UADFV Deep fake detection
Learning Detection Marquardt efficiency is to be
algorithm improved further.

As presented in Table 2, different methods are used for 2D to 3D image conversion and vice
versa particularly based on deep learning.

Table 3: Image processing techniques used for 2D and 3D image conversion

Reference Approach Technique Algorithm Data set Limitation /

Future Scope
[3] Post-imaging 3D SMLM and 2D-to-3D Custom The method
mathematical Super- localization dataset needs to be
approaches resolution algorithm improved by
imaging considering
spatial
information.
[5] Learning based CNN and image ICP AFLW2000- The modus
approach processing algorithm 3D and operandi of the
AFLW- proposed system
LPFA is yet to be
improved for
better
performance.
[8] Image processing High Efficiency Motion Custom In the future,
Video Coding prediction dataset better video
algorithm compression
techniques are to
be introduced.
[9] Mesoscale Slicing Mapping Custom To improve
modeling technique meshing dataset scalability and
approach algorithm computational
efficiency.
[12] Image- to- image Image fusion, DL Celebi In the future,
translation classification algorithm dataset GAN architecture
is to be explored.
[13] Color image Modular Chaotic Custom In the future, the
encryption Arithmetic and color image dataset 3D modular
approach Keyspace encryption chaotic map
algorithm needs be to
extended for
better
performance.

As presented in Table 3, there are different kinds of image processing methods used for 2D
to 3D image conversion besides providing their merits and demerits.

8
ITM Web of Conferences 64, 01010 (2024) https://doi.org/10.1051/itmconf/20246401010
ICACS24

5 Conclusion
A systematic review is made in this paper on existing methods of 2D to 3D image conversion and vice
versa. 3D object reconstruction from 2D images is a very active and useful field that has many practical
applications such as medical, military, security, logistics, and virtual tourism.
There are many more application fields for 3D object reconstruction such as Free-viewpoint video
reconstruction, robotic mapping, city planning, Gaming, virtual environments, robot navigation,
archaeology, augmented reality, reverse engineering, motion capture, Gesture recognition, and hand
tracking. Computer vision applications need image processing and learning-based techniques for
dealing with image content. In many computer vision applications, it is essential to deal with 2D and
3D images. It is also important to convert 2D to 3D and vice versa for different applications. This
systematic review throws light on the current level of innovation in image conversion techniques. Many
image processing techniques and deep learning models are found in the literature for image conversion.
CNN and CNN variants are widely used for image conversion. Transfer learning is used to enhance the
reuse of trained models. GAN-based approaches are also found useful for image conversion. Moreover,
different computer vision applications are needed to convert from either 2D to 3D or vice versa. Many
image processing techniques are also used for image conversion. From the insights of this paper, it is
observed that image conversion with learning-based approaches could leverage quality in conversion
besides enabling underlying applications for better performance.

References

1. Aharchi, M. and Kbir, A., A Review on 3D Reconstruction Techniques from 2D Images,

Innovations in Smart Cities Applications Edition 3 (pp.510-522),http://
DOI:10.1007/978-3-030-37629-1_37 , (2020)
2. Bostanabad, R. (2020). Reconstruction of 3D microstructures from 2D images via transfer
learning. Elsevier, 102906. http://doi:10.1016/j.cad.2020.102906
3. Broumandnia, A. The 3D modular chaotic map to digital color image encryption. Future
Generation Computer Systems, 99, pp. 489–499.
http://doi:10.1016/j.future.2019.04.005 , (2019).
4. Cai, W., Liu, D., Ning, X., Wang, C. and Xie, G. Voxel-based three-view hybrid parallel
network for 3D object classification. Elsevier. 69, pp.1-8.
https://doi.org/10.1016/j.displa.2021.102076 , (2021).
5. Eschweiler, D., Spina, Th., V., Choudhury, R. C., Meyerowitz E., Cunha, A. and
Stegmaier, J. CNN-BASED PREPROCESSING TO OPTIMIZE WATERSHED-
BASED CELL SEGMENTATION IN 3D CONFOCAL MICROSCOPY IMAGES.
IEEE., pp.1-5, (2019).
6. Feng, J., Teng, Q., Li, B., He, X. , Chen, H. and Li, Y. An end-to-end three-dimensional
reconstruction framework of porous media from a single two-dimensional image based
on deep learning. Elsevier, 368, 113043. http://doi:10.1016/j.cma.2020.113043, (2020).
7. Feng, Sh., Zuo, Ch., Yin, W., Gu, G. and Chen,Q. Micro deep learning profilometry for
high-speed 3D surface imaging. Optics and Lasers in Engineering, 121, 416–427.
http://doi:10.1016/j.optlaseng.2019.04.020 , (2019).
8. Junod, S., L., Tingey, M., Rush, C., Alkurdi, A. and Bajoria, K. Obtaining 3D super-
resolution images by utilizing rotationally symmetric structures and 2D-to-3D
transformation. Elsevier. 21, p.1424–1432. https://doi.org/10.1016/j.csbj.2023.02.008 ,
(2023).
9. Li, L., Lia, Z., Zakharchenko, V., Chen, J. and Li, H. Advanced 3D Motion Prediction
for Video Based Dynamic Point Cloud Compression. IEEE, pp.1–14.
http://doi:10.1109/TIP.2019.2931621, (2019).

9
ITM Web of Conferences 64, 01010 (2024) https://doi.org/10.1051/itmconf/20246401010
ICACS24

10. Li, M., Lin, Zh., Mech, R., Yumer, E. and Ramanan, D. Photo-Sketching: Inferring
Contour Drawings From Images. IEEE., pp.1-10.
http://DOI:10.1109/WACV.2019.00154, (2019).
11. Shamsolmoali, P., Zareapoor, M., Granger, E., Zhou, H., Wang, R., Celebi M. E. and
Yang, J. Image synthesis with adversarial networks: A comprehensive survey and case
studies. Elsevier. http://doi:10.1016/j.inffus.2021.02.014, (2021).
12. Tu, X., Zhao, J., Xie, M., Jiang, Z. , Balamurugan, K., Luo, Y., Zhao, Y., He, L. , Ma,
Zh. and Feng, J. 3D Face Reconstruction From A Single Image Assisted by 2D Face
Images in the Wild. IEEE, 23, pp. 1160- 1172. http://doi:10.1109/tmm.2020.2993962,
(2021).
13. Wu, O., Li, X., Wang, K. and Bilal, H. Regional feature fusion for on-road detection of
objects using camera and 3D-LiDAR in high-speed autonomous vehicles. Springer,
p.18195–18213. https://doi.org/10.1007/s00500-023-09278-3, (2023).
14. Wu, Zh., Zhang, J., Fang, Q., Yu, H. and Haiyan, M. Mesoscopic modelling of concrete
material under static and dynamic loadings: A review. Elsevier.
http://doi:10.1016/j.conbuildmat.2021.122419 , (2021).
15. Xu, Y., Tong, X. and Stilla, U. Voxel-based representation of 3D point clouds: Methods,
applications, and its potential use in the construction industry. Elsevier.
http://doi:10.1016/j.autcon.2021.103675, (2021).
16. Yang, X., Li, Y. and Lyu, S.. EXPOSING DEEP FAKES USING INCONSISTENT
HEAD POSES. IEEE., pp.8261-8265, (2019).

10