Review of DL-Based Generation Techniques of

Augmented Images using Portraits Specification

Mohammed Berrahal Mostafa Azizi
MATSI Research Lab MATSI Research Lab
ESTO, Mohammed First University ESTO, Mohammed First University
Oujda, Morocco Oujda, Morocco
[Link]@[Link] [Link]@[Link]

Abstract— Artificial Intelligence (AI) reached various With the advent in technology, many software is using
domains in our daily problems, among them is aiding law generate composite sketches, as such law enforcement has
enforcement in identifying suspects by generating face images started using these technologies. The eyewitness describes
or retrieving their images from existing databases, based on the
the portrait of the offender, in other hands the police officer
description of witnesses. The existing methods like hand-drawn
sketches take time and absorb human resources; in addition to
tries to sketch the face using the software such as EvoFIT [3],
that rendering images from the software fails most of the time Faces [4] and IdentiKit [5]. As an example, fig.1 (a) is a
to sketch a real image and does not get the ideal scenario. In this digital image next to it (b) it’s a hand-drawn face based on an
paper, we present a survey of the recent progress concerning eyewitness, in (c) the image is corresponding computer-
generating images from description and sketch face recognition, generated using IdentiKit, finally (d) images generated using
we analyze the difference between several algorithms on the FACES [4].
problem based on evaluation metrics like accuracy and
correlation similarity. We also give an overview of datasets used The rendering images from the software fails most of the
for generations or recognition. time to sketch a real image and does not get the ideal scenario
[7]. End duo the limitations of viewed sketches and
Keywords— Generative Adversarial Networks, Composite emphasized the need for new databases and algorithms
Face, Sketch Face, Deep Learning, Image Generation. imitating real scenarios, without forgetting the waste of time
I. INTRODUCTION and human resources, especially if time is important in the
investigation.
To solve serious murder investigations, crime rape and
burglary police officers try to determine the person involved The rest of this paper is organized as follows: In the second
using facial composite, from eyewitnesses’ memories, this section, we recall backgrounds about facial composite,
technique is known as a graphical representation of the face artificial intelligence (AI), Machines Learning (ML), Deep
Learning (DL), Generative Adversarial Networks (GAN) and
drawn by a forensic technician or an advanced program.
Autoencoder (AE). In the third section, we present our
These images are used to reconstruct the suspect's face, the
research methodology of literature review. Before concluding,
resulting image is circulated within a member of the the fourth section discusses our findings about the different
investigation authority or in the media: TV, newspapers, and techniques of image generation using DL concepts.
social media for identification, to locate the offender or
provide a lead to the investigation [1][2].

Fig. 1. Comparison of hand-drawn sketches and composite sketches

978-1-7281-8084-7/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 02,2021 at [Link] UTC from IEEE Xplore. Restrictions apply.
 can be categorized in three parts: supervised learning, semi-
supervised learning or unsupervised learning. Work with
II. TERMINOLOGIES multiple layers for the extraction of features from input raw.
A. Facial Composite The structure of a neural network resembles the structure and
Facial Composite is the art of drawing or sketching the function of the human brain, with layers of connected nodes
human face based on the description of the eyewitnesses, that can learn from data and be trained to recognize patterns,
thanks to the variety of programs, this method is no longer classify data, and predict events. Deep learning models can
dependent on the artist skills, these programs provide drag learn from large datasets, this architecture of the models learn
and drop options for separate facial features, to produce the directly from the data without the requirement for extraction.
composite image [1]. This method is capable of D. Generative Adversarial Networks
reconstructing a true face image for identifying or verifying a For generating an unsupervised data, multiple ML and DL
person. based algorithms have been developed. A very powerful way
So far, incredible achievement had been made to of doing this is utilizing generative adversarial networks [5]
elaborate algorithms and methods for composite faces based also known as GAN.
on deep learning [2], especially with the success of The idea behind GANs as shown in fig.2 is the
Generative Adversarial Networks (GANs), like the example combination of two neural networks, a generator G and the
of composition-aided generative adversarial network (CA- discriminator D. the first G gets a randomly sampled noise
GAN) for face photo-sketch synthesis [3]. In the near future, vector as an input; in most cases we sample from a Gaussian
this technique will be implemented on machines, without distribution a noise vector to feed the neural network with a
human interaction. bunch of convolutional layers. We at the end an image that
B. Artificial Intelligence (AI) we direct to D. and the discriminator needs to decide whether
Artificial intelligence is a combination of concepts and that image comes from the actual dataset (the real images that
technologies that recreate or at least imitate the intelligence were training on) or from the generator. We represent the
of humans, like self-driving cars, and autonomous robots that discriminator with the function D and the generator with the
impersonate humans and their patterns, including perception, G in the following equation:
reasoning, learning, communicating and acting in complex minG maxD V (D, G) = Ex~pdata(x)[logD(x)] + Ez~pz(z)[log(1-D(G(z))]. (1)
environments. The ultimate goal is the development of
machines that can perform tasks as well as humans and more where:
[4] – its use is spreading everywhere.
In our work we focus on the category Machine Learning
x D(x) is the discriminator's estimate of the probability that
(ML) with its subfield deep learning (DL).
real data instance x is real.
x Ex~pdata(x) is the expected value over all real data
C. Machines Learning (ML) and Deep Learning (DL)
instances.
ML Machine Learning a sub-area of AI is the ability
x G(z) is the generator's output when given noise z.
to find a solution indecently by recognizing patterns from
datasets, and improving the solution through the experience x D(G(z)) is the discriminator's estimate of the probability
via multiple algorithms like finding and extracting data from that a fake instance is real.
datasets, make predictions on data analytics and optimize the x Ez~pz(z) is the expected value over all random inputs to the
process of recognizing patterns. These algorithms are often generator (in effect, the expected value over all generated
categorized as supervised or unsupervised. fake instances G(z)).
a) The Supervised algorithms main’s function is to
predict future events using labeled data and based on what The generator can't directly affect the log(D(x)) term in
has been learned in previous cases. First they analyze training the function, so, for the generator, minimizing the loss is
dataset, and thanks to learning algorithms, they can make equivalent to minimizing log (1 - D(G(z))).
predictions about the output values by producing an inferred Whether that image was a real one coming from the
function. Then the system is able to provide targets for any dataset or a fake image issued from the generator, and by
new input after sufficient training. These type of algorithms using this label we can basically back propagate a training
can find errors by comparing its output with the correct output loss through this discriminator network in order to make it
to modify the model correspondingly. better. The generator network itself is also a fully
b) Unsupervised algorithms are used when the differentiable neural network so if two networks work back-
information used to train is neither classified nor labeled. to-back, we can actually back propagate the learning signal
Unsupervised learning studies how systems can infer a through this entire model pipeline. By this way we can update
function to describe a hidden structure from unlabeled data. with the same single loss function, both the discriminator and
The system doesn’t figure out the right output, but it explores the generator networks until they both get really good at their
the data and can draw inferences from datasets to describe jobs. So the most important trick in this pipeline is to make
hidden structures from unlabeled data. sure that both these networks are well balanced during
DL is considered a subfield of machine learning involving training so that no one of them gets the upper hand and if you
different algorithms structured to behave like a human’s manage to do so and you train it for long enough then
brain, called artificial neural networks, this learning method eventually what you'll have is a generator that has been

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 02,2021 at [Link] UTC from IEEE Xplore. Restrictions apply.
learning from the feedback of this discriminator network and
can eventually manage to generate images that look very
similar to the dataset.
Training the Generator Training the Discriminator Collecte real Images

Randome Noise Z Real Images

G
DATASET
FAKE REAL

Fake Sample LOSS Real Sample

Fig. 2. Illustrating the Generative Adversarial Model

Ideally they are identical Reconstructed

Input
Input
Latent space
Representation

x Z x'

Encoder Bottleneck Decoder

Fig. 3. Schema of an Autoencoder.

decoding network part, W is the weight matrix, and b is the

E. Autoencoder (AE)
bias vector.
AE is a king of unsupervised non-recurrent feed forward
neural networks. The structure of the Autoencoder is III. RESEARCH METHOD
constituted by an encoder that maps the input to a code, latent This review literature is performed to identify and analyse
variable or latent representation, in another hand the decoder research works describing virtual images through Deep
maps the code, latent variable or latent representation to the Learning based generation techniques of augmented images
input with the same structure. Finally, the layers are divided using portraits specification to deal with the problem of Sketch
into three-part input layer, output layer and one or multiple Based Face Recognition (SBFR). Studied papers were first
hidden layers. The principal technique used here is the back identified through manual searches over online documentary
propagation, the purpose is to train the model layer-by-layer databases (Google Scholar, Scopus, Web of Science,
by setting the target values to be equal to the inputs[6]. The ScienceDirect, Semantic Scholar, and IEEE Xplore) by using
neural is trained to minimize the reconstruction errors. The following keywords “Deep Learning,” “Face Sketch
fig 3. represents a basic model of an AE. Recognition” and “Generative Adversarial Networks Face
Sketch Recognition.”
The main function that represents this model is: A. Search and Selection
ଶ ଶ We regroup a total of 470 papers describing Face Sketch
ࣦሺ‫ݔ‬ǡ ‫ ݔ‬ᇱሻ ൌ หȁ‫ ݔ‬െ ‫ ݔ‬ᇱȁห ൌ  ቚห‫ ݔ‬െ ߪ ᇱ ሺܹ ᇱ ൫ߪሺܹ௫ ൅ ܾሻ൯ ൅ ܾԢหቚ (2) Recognition with Deep Learning related to computer vision.
Details of searched papers are included in Table.1 and fig.4.
‫ ݖ‬ൌ ߪሺܹ‫ ݔ‬൅ ܾሻis called the encoding network part or the Table 1 represents the total number of results from the web
latent dimension. Similarly, ‫ ݔ‬ᇱ ൌ ߪ ᇱ ሺܹ ᇱ ‫ ݖ‬൅ ܾ ᇱ ሻ is the library. fig.4 illustrates the number of papers in each year from

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 02,2021 at [Link] UTC from IEEE Xplore. Restrictions apply.
2012 to 2019. It can be seen that the number of papers Table 1. Results from searched databases
increases each year from 2012 to 2019. As our work written 2
May 2020, this number cannot represent the overall number Library Total number of results
of papers in 2020. Especially since there are several upcoming Scopus 432
top-tier conferences, recognition face sketch still attracts lots Web of Science 333
of attention and number of papers is increasing. ScienceDirect 15651
Google Scholar 597000
Semantic Scholar 159000

B. Synthesizing the literature

To compare existing algorithms for recognizing and
classifying sketch faces it requires a relatively large number
of images with multiple facial features, we regroup all
databases treated in our literature review, as shown in Table
2.

Fig. 4. Scopus, analyze search result of the search key, “Face

Sketch Recognition.

Table 2. Available face images and face sketch datasets

Ref Data name Type of images Number of Context

images
[7] CUHK Face Sketch Face 188 Data set contain face images, for each face, there is a hand drawn sketch
(Sketches, by an artist based on a photo taken to the face, under typical lighting
AR data set 123
Images) condition, and with a nonpartisan articulation.

XM2VTS 295
[8] (CelebA) Dataset Face images 202,599 The pictures in this dataset spread enormous posture varieties and
background clutter. CelebA has large diversities, large quantities,
and rich annotations, including.
[9] Labeled Faces in the Wild Face images 13,233 Database of face, designed for studying the problem of
(LFW) unconstrained face recognition, each face has been labelled with the
name of the person pictured.
[10] Face Recognition Face images 14,126 The FERET database contains 1564 sets of images for a total of
Technology: 14,126 images that includes 1199 individuals and 365 duplicate sets
Color FERET Database of images.

[11] Aberdeen DataSet 2D face sets 687 Aberdeen Dataset contains 687 color faces from, Between 1 and 18
of 90 individuals
[12] The Japanese Female Facial 213 The database contains 213 images of 7 facial expressions (6 basic
Facial Expression expressions facial expressions + 1 neutral) posed by 10 Japanese female
(JAFFE) Database models.

[13] The Radboud Faces Facial 536 Database RaFD contains 536 images in total distributed as 67
Database (RaFD) expressions people displaying 8 emotional expressions.
[13] Faces95 Face images 1,440 Database Faces95 contains 1440 images in total 72 Number of
individuals taken 20 per person.
[14] The CMU Multi-PIE Face Face images More This database contains in total more than 750,000 face images of
Database than 337 people the size of the data surpasses 305 GB.
750,000
[15] The Pattern Recognition Face 265 The dataset contains hand-draw composites based on the verbal
and Image Processing (Sketches, description by the eyewitness.
(PRIP) Hand-Drawn Images)
Composite (PRIP-HDC)

Table 3. Summary of reviewed literature.

Paper Title Method Approach Dataset Metric

[3] Towards Realistic Face Photo- Composition- This approach improves the realism of the synthesized face [7] from 89% to
Sketch Synthesis via Aided GANs photos and sketches by proposing a compositional 98.8%
Composition-Aided GANs reconstruction loss and using stacked CA-GANs.

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 02,2021 at [Link] UTC from IEEE Xplore. Restrictions apply.
Paper Title Method Approach Dataset Metric
[16] Coupled Generative Adversarial CoGAN Unsupervised domain adaptation and image transformation RGBD / 72%
Networks tasks NYU

[17] Face Generation with Conditional CGAN Face generation conditioned on identity by implementing [8] 98% detection
Generative Adversarial Networks variants with multiple network architectures and loss rate and 60%
functions accuracy

[18] Identity-Aware CycleGAN for CycleGAN Jointly solve synthesis and recognition problem by their close [7] Recognition
Face Photo-Sketch Synthesis and relationship accuracy
Recognition approaches
95.76%
[19] Facial attribute-controlled sketch- GAN - sketch- High-quality sketch-to-image translation control the facial [8] Image Quality
to-image translation with to-image attributes of synthesized images
generative adversarial networks translation
[20] Generative Adversarial Network SaGAN Change the attribute-irrelevant regions [8] - [9] Image Quality
with Spatial Attention for Face
Attribute Editing
[21] Generative Adversarial Networks GAN & The purpose is improving the results of an existing facial [9] Accuracy of
for Improving Face Classification Softmax recognition approach by utilizing generative adversarial 99.42%
GAN networks to improve the existing dataset.
[22] Local Deep Features for local deep This work uses two deep models DEEPS (trained for sketch [10] - [15] RECOGNITION
Composite Face Sketch learning recognition) and ResNet-Dlib (trained for face recognition). It RATE 76.3%
Recognition was shown that in general the proposal using the local features
from ResNet-Dlib model achieves the highest recognition rates
[23] Person‑independent facial Wasserstein This method based on Wasserstein generative adversarial [12] - [24] 96% accuracy
expression recognition method generative networks for micro-facial expressions established to improve
based on improved Wasserstein adversarial the accuracy and robustness of outward appearance
generative adversarial networks networks acknowledgment by means of hindrance of intra-class variety.
in combination (WGAN)
with identity aware
[25] StarGAN: Unified Generative StarGAN Image-to-image translations for multiple domains using only a [8] - [13] 99.55%
Adversarial Networks single model accuracy
for Multi-Domain Image-to-
Image Translation
[9] Towards Large-Pose Face FF-GAN The framework differs from both traditional GANs and 3DMM [9] - [24] 96.42%
Frontalization in the Wild based modelling. Incorporating 3DMM into the GAN structure accuracy
provides shape and appearance priors for fast convergence with
less training data
[26] APDrawingGAN: Generating APDrawingG The purpose of this neural network is to transform a face photo APDrawin 62.14 FID
Artistic Portrait Drawings AN into an APDrawing by learning all drawing strategies for g dataset
from Face Photos with different facial features, and develop loss to measure similarity
Hierarchical GANs between generated and artists.
[27] Automated Criminal GAN & HOG Generation of face images based on image generation by using [8] - [9] 99.63%
Identification System Algorithm their sketch and visual attributes accuracy
Using Face Generation
[28] Evaluating the Performance of CGAN & CNN Methodology for generating a colored photo from a sketch, to [7] - [8] - 65%
Face Sketch Generation using identification using a variety of classification techniques. [15] similarity
Generative Adversarial 98% accuracy
Networks
[29] Generative Adversarial Multitask Generative Improved deep network named residual dense U-Net, multitask [7] CUFS: 96%
Learning for Face Sketch Adversarial discriminator to guide the accuracy -
Synthesis and Recognition Multitask generator to produce more realistic sketch image, and extract CUFSF: 98%
learning discriminative face feature. accuracy

[30] Human Sketch Recognition using CGAN & Sketch recognition model using multi-layered approach by [8] - [9] - 97.2%
Generative Adversarial Networks FaceNet combining sketch to image generation and face recognition [11] - [14] accuracy
and One-Shot Learning methods
[31] Integrating GAN with CNN for GAN & CNN Applying deep [7] - [10] 82.85%
Face Sketch Synthesis learning features by combining with GAN and CNN model accuracy
ResNet to generate synthesize images
[32] Synthesis of Facial Image using CGAN This method uses CGAN to generate realistic images that [7] FR
Conditional preserve the identity the target person verified by face score
Generative Adversarial Network recognition algorithms 96.23%
[33] Composite Sketch Recognition Deep Network The proposed algorithm performs inductive transfers on the [5] 58%
via Deep Network - A Transfer features learned using a deep learning architecture to effectively accuracy
Learning Approach match the heterogeneous information.

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 02,2021 at [Link] UTC from IEEE Xplore. Restrictions apply.
 Among all the papers published related to our literature ே௨௠௕௘௥௢௙஼௢௥௥௘௖௧௣௥௘ௗ௜௖௧௜௢௡௦
‫ ݕܿܽݎݑܿܿܣ‬ൌ  ்௢௧௔௟௡௨௠௕௘௥௢௙௣௥௘ௗ௜௖௧௜௢௡௦௠௔ௗ௘ (3)
review we singled out those between 2012 and 2020, it turned
out to be 19 articles in total we analysed and inspected all the
data presented and the results given and integrated it in Table Other methods suggested to compare correlation
3, as shown. similarity for more realistic images they use correlation. The
similarity depends on how much two items are correlated, the
IV. FINDINGS higher the correlation is the higher the similarities, as
We found ten datasets used to analyze SBFR algorithms, represented in the following formula [35] :
some of them contain two types of images sketches and
ஊሺ୶ି௫ҧ ሻሺ௬ି௬തሻ
photos. According to these images, we can distinguish 4 types ‫ ݎ‬ൌ (4)
ඥஊሺ୶ି௫ҧ ሻమ ஊሺ௬ି௬തሻమ
of sketch face recognition methods hand-drawn viewed
sketches, hand-drawn semi-forensic sketches, hand-drawn
Also, there are other methods like Detection Rate, Fréchet
forensic sketches and software-generated composite sketches
Inception Distance and Face Recognition Score, it depends on
photos. certain algorithms. Some literatures just compare visually the
quality of the image with previous existing states of art, we try
In Our work, we present different approaches of BSFR in fig.5 to regroup some of the potential papers, using
and generation images, we use the accuracy to compare Classification Tree Method (CTM) to compare them.
between them. The classification accuracy is the ratio of the
numbers of correct predictions to the total number of input
sample[34].

DATASET

(CUHK) (CELEB A) (LFW) (FERET) (PRIP)

[3] [28] [18] [30] [32] [31] [27] [23]

GAN
DL USED ALGORITHMS

CoGAN

CGAN

CNN

CycleGAN

WGAN

98.8% 98% 60% 97.2% 82.85% 99.63%

Accuracy
COMPARISON METHODS

98% 76.3%
Detection Rate

Correlation 65%
Similarity

Face 96.23%
Recognition
Score

Fig. 5. Combination use of datasets and DL techniques over some studies papers.

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 02,2021 at [Link] UTC from IEEE Xplore. Restrictions apply.
 The highest accuracy reached for generation sketch 10.5281/ZENODO.3451524.
[13] O. Langner et al., “Presentation and validation of the Radboud
images was found on CUHK datasets mostly with variation Faces Database,” doi: 10.1080/02699930903485076.
of GAN algorithms, or combination of GAN with another [14] “Face Recognition Data.” [Online]. Available:
algorithm. The same thing with generation real images [Link] [Accessed:
happens with CelebA and LFW datasets due to diversity and 07-May-2020].
[15] “The Pattern Recognition and Image Processing (PRIP) Hand-
the number of images that are available in these datasets. Drawn Composite (PRIP-HDC).” [Online]. Available:
Even if this area shows promising results, the image [Link]
generation form descriptions of eyewitnesses are still HDC-Release_ReadMe.txt. [Accessed: 07-May-2020].
challenging than the classical face generation. The main [16] M.-Y. Liu and O. Tuzel, “Coupled Generative Adversarial
reason is the lack of large databases and the heterogeneous Networks,” [Link], no. Nips, Jun. 2016.
nature of sketch and photo modalities, the methods that we [17] X. Cao, S. Rao Dulloor, and M. C. Prasetio, “Face Generation with
Conditional Generative Adversarial Networks,” 2017.
have seen try to reduce the lack between the two-area motions. [18] Y. Fang, W. Deng, J. Du, and J. Hu, “Identity-aware CycleGAN
Also, the sketch images do not comport some face features for face photo-sketch synthesis and recognition,” Pattern
and characteristics with wearing sunglasses or hate. Every Recognit., vol. 102, p. 107249, Jun. 2020, doi:
additional information can be effective to increase our 10.1016/[Link].2020.107249.
accuracy and determine the offender [6]. [19] M. Hu and J. Guo, “Facial attribute-controlled sketch-to-image
translation with generative adversarial networks,” Eurasip J.
Image Video Process., vol. 2020, no. 1, pp. 1–13, Dec. 2020, doi:
V. CONCLUSION 10.1186/s13640-020-0489-5.
This paper outlines Deep Learning Based methods for [20] G. Zhang, M. Kan, S. Shan, and X. Chen, “Generative adversarial
Face Sketch Recognition (BFSR) purposes by transforming network with spatial attention for face attribute editing,” in Lecture
Notes in Computer Science (including subseries Lecture Notes in
sketches on real images using variety of algorithms such as Artificial Intelligence and Lecture Notes in Bioinformatics), 2018,
GAN, AE, CNN and transfer learning. vol. 11210 LNCS, pp. 422–437, doi: 10.1007/978-3-030-01231-
Most of the papers that we analysed try to generate facial 1_26.
photos or convert sketches to photos. Feature descriptor [21] J. N. Supervisor and M. Goodwin, “Generative Adversarial
Networks for Improving Face Classification,” 2017.
methodology for extracting the differential features from the [22] H. Mendez-Vazquez, F. Becerra-Riera, A. Morales-Gonzalez, L.
images is studied. Feature-based approaches outperform the Lopez-Avila, and M. Tistarelli, “Local deep features for composite
conversion function since the transformation method needs face sketch recognition,” in 2019 7th International Workshop on
huge memory as well as more time and also it depends on Biometrics and Forensics, IWBF 2019, 2019, doi:
training set of photos – sketches correspondence. Accuracy of 10.1109/IWBF.2019.8739212.
the system depends on sketch levels, so there is a need for [23] C. Xu, Y. Cui, Y. Zhang, P. Gao, and J. Xu, “Person-independent
facial expression recognition method based on improved
large collection of forensic sketches to further understand the Wasserstein generative adversarial networks in combination with
complexity of the problem. identity aware,” in Multimedia Systems, 2020, vol. 26, no. 1, pp.
53–61, doi: 10.1007/s00530-019-00628-6.
REFERENCES [24] “The CMU Multi-PIE Face Database.” [Online]. Available:
[1] R. C. L. Lindsay, Handbook Of Eyewitness Psychology 2 Volume [Link]
Set. Routledge, 2014. Pie/[Link]. [Accessed: 07-May-2020].
[2] M. Zhang, N. Wang, X. Gao, and Y. Li, “Markov Random Neural [25] Y. Choi, M. Choi, M. Kim, J. W. Ha, S. Kim, and J. Choo,
Fields for Face Sketch Synthesis,” 2017. “StarGAN: Unified Generative Adversarial Networks for Multi-
[3] J. Yu et al., “Towards Realistic Face Photo-Sketch Synthesis via domain Image-to-Image Translation,” in Proceedings of the IEEE
Composition-Aided GANs,” Dec. 2017. Computer Society Conference on Computer Vision and Pattern
[4] “Artificial Intelligence: A New Synthesis - Nils J. Nilsson, Nils Recognition, 2018, pp. 8789–8797, doi:
Johan Nilsson - Google Livres.” [Online]. Available: 10.1109/CVPR.2018.00916.
[Link] [26] R. Yi, Y. J. Liu, Y. K. Lai, and P. L. Rosin, “Apdrawinggan:
C&oi=fnd&pg=PP2&dq=artificial+intelligence&ots=Q_cqYM2y Generating artistic portrait drawings from face photos with
eF&sig=drgX5mhlf1brvtF_oiwdINj7F88&redir_esc=y#v=onepa hierarchical gans,” in Proceedings of the IEEE Computer Society
ge&q&f=false. [Accessed: 04-May-2020]. Conference on Computer Vision and Pattern Recognition, 2019,
[5] I. J. Goodfellow et al., “Generative adversarial nets,” in Advances vol. 2019-June, pp. 10735–10744, doi:
in Neural Information Processing Systems, 2014, doi: 10.1109/CVPR.2019.01100.
10.3156/jsoft.29.5_177_2. [27] A. Karve, M. Balasubramanian, K. Chaudhari, and S. B. Mane,
[6] G. Guo and N. Zhang, “A survey on deep learning based face “Automated Criminal Identification System Using Face
recognition,” Comput. Vis. Image Underst., vol. 189, p. 102805, Generation,” in Advances in Intelligent Systems and Computing,
Dec. 2019, doi: 10.1016/[Link].2019.102805. 2020, vol. 1039, pp. 566–573, doi: 10.1007/978-3-030-30465-
[7] “CUHK Face Sketch Database (CUFS).” [Online]. Available: 2_63.
[Link] [Accessed: [28] M. S. Sannidhan, G. Ananth Prabhu, D. E. Robbins, and C. Shasky,
07-May-2020]. “Evaluating the performance of face sketch generation using
[8] “Large-scale CelebFaces Attributes (CelebA) Dataset.” [Online]. generative adversarial networks,” Pattern Recognit. Lett., vol. 128,
Available: [Link] pp. 452–458, Dec. 2019, doi: 10.1016/[Link].2019.10.010.
[Accessed: 07-May-2020]. [29] W. Wan and H. J. Lee, “Generative Adversarial Multi-Task
[9] X. Yin, X. Yu, K. Sohn, X. Liu, and M. Chandraker, “Towards Learning for Face Sketch Synthesis and Recognition,” in
Large-Pose Face Frontalization in the Wild,” in Proceedings of the Proceedings - International Conference on Image Processing,
IEEE International Conference on Computer Vision, 2017, vol. ICIP, 2019, vol. 2019-September, pp. 4065–4069, doi:
2017-October, pp. 4010–4019, doi: 10.1109/ICCV.2017.430. 10.1109/ICIP.2019.8803617.
[10] “Face Recognition Technology (FERET) | NIST.” [Online]. [30] D. Wadhwa, U. Maharana, D. Shah, V. Yadav, and P. Pandey,
Available: [Link] “Human Sketch Recognition using Generative Adversarial
recognition-technology-feret. [Accessed: 07-May-2020]. Networks and One-Shot Learning,” in 2019 12th International
[11] “2D face sets.” [Online]. Available: Conference on Contemporary Computing, IC3 2019, 2019, doi:
[Link] [Accessed: 07-May- 10.1109/IC3.2019.8844885.
2020]. [31] A. Akram, N. Wang, X. Gao, and J. Li, “Integrating GAN with
[12] M. Lyons, M. Kamachi, and J. Gyoba, “The Japanese Female CNN for face sketch synthesis,” in 2018 IEEE 4th International
Facial Expression (JAFFE) Database,” Apr. 1998, doi: Conference on Computer and Communications, ICCC 2018, 2018,

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 02,2021 at [Link] UTC from IEEE Xplore. Restrictions apply.
 pp. 1483–1487, doi: 10.1109/CompComm.2018.8780648.
[32] S. Roy, M. A. H. Akhand, and N. Siddique, “Synthesis of Facial
Image using Conditional Generative Adversarial Network,” in 5th
International Conference on Computer, Communication,
Chemical, Materials and Electronic Engineering, IC4ME2 2019,
2019, doi: 10.1109/IC4ME247184.2019.9036488.
[33] P. Mittal, M. Vatsa, and R. Singh, “Composite sketch recognition
via deep network - A transfer learning approach,” in Proceedings
of 2015 International Conference on Biometrics, ICB 2015, 2015,
pp. 251–256, doi: 10.1109/ICB.2015.7139092.
[34] “Metrics to Evaluate your Machine Learning Algorithm.”
[Online]. Available: [Link]
evaluate-your-machine-learning-algorithm-f10ba6e38234.
[Accessed: 01-Jun-2020].
[35] “Similarity and Distance Metrics for Data Science and Machine
Learning.” [Online]. Available:
[Link]
for-data-science-and-machine-learning-e5121b3956f8.
[Accessed: 01-Jun-2020].

Authorized licensed use limited to: University of Prince Edward Island. Downloaded on June 02,2021 at [Link] UTC from IEEE Xplore. Restrictions apply.