156 IEEE SIGNAL PROCESSING LETTERS, VOL.

27, 2019

Unsupervised Face Domain Transfer for

Low-Resolution Face Recognition
Sungeun Hong , Member, IEEE, and Jongbin Ryu , Member, IEEE

Abstract—Low-resolution face recognition suffers from domain

shift due to the different resolution between a high-resolution
gallery and a low-resolution probe set. Conventional methods
use the pairwise correlation between high-resolution and low-
resolution for the same subject, which requires label information
for both gallery and probe sets. However, explicitly labeled low-
resolution probe images are seldom available, and labeling them
is labor-intensive. In this paper, we propose a novel unsupervised
face domain transfer for robust low-resolution face recognition.
By leveraging the attention mechanism, the proposed generative
face augmentation reduces the domain shift at image-level, while
spatial resolution adaptation generates domain-invariant and dis-
criminant feature distributions. On public datasets, we demon- Fig. 1. Overview of the proposed method. The gallery set contains labeled
strate the complementarity between generative face augmentation high-resolution (HR) images, while the probe set includes low-resolution (LR)
at image-level and spatial resolution adaptation at feature-level. images without labels. Subscript ‘1’ or ‘2’ in the legend refers to the class label.
The proposed method outperforms the state-of-the-art supervised Given HR images, generative face augmentation (GFA) synthesizes ‘HR-LR’
methods even though we do not use any label information of images so that they resemble LR images at the image-level. Spatial resolu-
low-resolution probe set. tion adaptation (SRA) enables common feature spaces in both domains to be
discriminative.
Index Terms—Low-resolution face recognition, image-to-image
translation, domain adaptation, attention, face augmentation.
sets over the past decades. Among these studies, there has
been an increasing interest in low-resolution face recognition
I. INTRODUCTION (LRFR) [14], which takes probe images under low-quality con-
LTHOUGH recent face recognition methods have shown ditions [15]–[17]. The existing methods can be divided into
A promising results [1]–[3], heterogeneous domains be-
tween the gallery and probe sets still remains a challenging
two categories. Hallucination approaches obtain high-resolution
(HR) images from low-resolution (LR) probe images before per-
problem [4], [5]. A popular example of the heterogeneity is the forming face recognition [18]–[20]. On the other hand, models
significant difference between a high-quality gallery (e.g., iden- in the embedding category generate common feature space for
tification photos) and low-quality probe images from surveil- HR and LR conditions [21]–[23]. Crucially, both approaches use
lance cameras [6]. Photos used in ID cards or passports are identity labels for both HR and LR images. In other words, most
captured in a stable environment to register as gallery images. existing LRFR methods learn the relations between HR gallery
On the other hand, probe images are usually captured under and LR probe sets through the supervision of identity labels.
unstable real environments including noise, blur, arbitrary pose However, we argue that unsupervised approaches that do not
as mentioned in [7], [8]. This causes severe domain shift between require the labeling of LR probe sets have significant advantages
source (i.e., gallery) and target (i.e., probe) images, which de- over supervised-based approaches. First of all, while it is not
grades the recognition performance. difficult to acquire LR images in advance, explicitly labeled
Considerable efforts [9]–[13] have been devoted to coping images are seldom available and labeling them is quite labor-
with the heterogeneous condition between the gallery and probe intensive and impractical. Second, the unsupervised setting has
the advantage to improve the performance of the recognition
model over time. Unsupervised face recognition systems do not
Manuscript received October 15, 2019; revised December 11, 2019; accepted require any label information of LR probe images, so they can
December 19, 2019. Date of publication December 30, 2019; date of current
version January 31, 2020. This work was supported by Basic Science Research easily evolve even the LR probe images are continuously given.
Program through the National Research Foundation of Korea(NRF) funded In this paper, we propose an unsupervised LRFR method
by the Ministry of Education (NRF-2017R1A6A3A11031193). The associate to reduce domain shift at feature-level as well as image-level.
editor coordinating the review of this manuscript and approving it for publication
was Dr. S. S. Channappayya. (Corresponding author: Jongbin Ryu.) Unlike previous approaches [18], [19], [21]–[23], our model
S. Hong is with T-Brain, AI Center, SK Telecom, 04539 Seoul, South Korea uses only images without label information in an LR probe
(e-mail: [email protected]). set. As can be seen in Fig. 1, the proposed method con-
J. Ryu is with the Department of Computer Science, Hanyang University,
04763 Seoul, South Korea (e-mail: [email protected]). sists of generative face augmentation (GFA) and spatial res-
Digital Object Identifier 10.1109/LSP.2019.2963001 olution adaptation (SRA) networks to produce a discriminant
1070-9908 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: Odisha University of Technology and Research. Downloaded on October 24,2024 at 08:10:04 UTC from IEEE Xplore. Restrictions apply.
 HONG AND RYU: UNSUPERVISED FACE DOMAIN TRANSFER FOR LOW-RESOLUTION FACE RECOGNITION 157

low-resolution facial images. Hallucination-based reconstruc-

tion of HR images from input LR images was also studied.
To transfer LR images into HR images, super-resolution tech-
niques or facial attribute embedding methods have been widely
used [11], [20], [32]. These methods transfer probe LR images
into HR images, which cause over smoothness and loss of
details. Conversely, we transfer HR images into LR images by
using attention-guided domain transfer and also perform domain
alignment at feature-level.
On the other hand, the goal of the feature-level category is to
find a unified feature space where the proximity between LR and
Fig. 2. Given labeled HR images and unlabeled LR images, our GFA can HR images is maintained for the same subject. Coupled mapping
map two domains without any pair labels by leveraging attention mechanism. strategy is widely used for learning the projection between HR
(a) (from left to right) HR gallery images, synthesized images with varying images and LR images [14], [21]. Dictionary-based approaches
pose, learned attention map in which bright value indicates higher focus, and
domain-transferred HR-LR images with various poses (b) LR probe images. have also been suggested to match facial images captured at
different resolutions [22], [23]. Crucially, most existing methods
assume that there are both LR and HR versions available for
distribution that is aligned at both image and feature levels. The each subject, which is impractical for real-world applications.
proposed GFA is designed to synthesize labeled HR images into In contrast, we propose an unsupervised method that does not
LR-like images while SRA ensures that generated LR-like im- require any labels in LR images.
ages are indistinguishable from real LR images at feature-level. Domain Adaptation (DA): Several attempts have been pro-
The proposed method is closely related to our previous posed to apply domain adaptation (DA) to face-related tasks.
work [24] in considering domain adaptation networks for het- Xie et al. [33] use DA and several hand-crafted descriptors to
erogeneous face recognition. In our prior work [24], we generate tackle face recognition in which the gallery set consists of clear
synthetic images with a 3D face model, focusing only on pose images while the probe set includes blurred images. Banerjee
variation. In contrast, this work introduces joint optimization of et al. [34] propose a DA-based method with a filter bank,
the learning-based generative network (i.e., GFA) with domain e.g., Eigenfaces, Fisherfaces, and Gaborfaces. Unlike the above
adaptation network (i.e., SRA) to generate domain-transferred approaches, which apply DA after extracting the handcrafted-
images and also domain-invariant feature distribution. However, features from images, we jointly perform feature learning and
simply transferring images from HR domain to LR domain classification in an integrated deep architecture. Moreover, we
frequently results in mode collapse [25] and unexpected arti- solve the face recognition problem in which only one labeled
facts [26]. Therefore, our GFA includes the attention-guided image per person is given to the model, which is a challenging
domain transfer module to focus on critical parts for compen- and realistic protocol.
sating the difference between the source and target images. Fig. 2
shows sample images in HR gallery images, domain-transferred III. PROPOSED METHOD
images, and LR probe images. A. Generative Face Augmentation (GFA)
We validate the proposed method in a challenging and realistic
LRFR protocol on public datasets [24], [27], [28]. The gallery set The goal of GFA is to learn a mapping function from source
includes HR images captured under stable environments, such HR domain to target LR domain, i.e., GenHR→LR . To bridge the
as identification photos, while the probe set contains unlabeled gap between them, we use image-level domain transfers using
LR images from surveillance cameras. We experimentally show ‘pix2pix’ framework [35], which consists of two sub-networks:
that our proposed GFA and SRA perform complementarily 1) auto-encoder architecture called ‘generator’ that attempts to
and eventually show drastic performance improvement. Sur- translate an image 2) ‘discriminator’ trying to classify whether
prisingly, the accuracy of our unsupervised model is superior the image is actually translated from ‘generator.’ Due to this
to the state-of-the-art supervised methods [10], [24], [29]–[31]. generative adversarial relationship between them as in gen-
We also provide extensive ablation studies that can serve as erative adversarial networks (GAN) [36], both generator and
quantitative baselines for domain adaptation-based approaches discriminator are getting better as learning progresses.
for LRFR. Since pair data of HR and LR images for the same subject
is not available, we construct two generators and adopt a cycle
consistency loss [37]. However, simply applying image-level
II. RELATED WORKS domain transfer to the source and target domains can lead to
Low-Resolution Face Recognition (LRFR): Conventional unexpected artifacts and performance degradation. Motivated
LRFR methods can be grouped into two categories depending by the recent success of attention-based approaches in computer
on the level at which resolution alignment is applied, i.e., vision fields [38]–[40], we leverage attention mechanism during
image-level or feature-level. Early investigations in image- a domain transfer. To this end, we add two attention modules
level methods include singular value decomposition [18] or AttHR−>LR and AttLR−>HR , which reduce the artifacts of
sparse representation [19] for synthesizing person-specific synthesized images by focusing only on crucial parts in the

Authorized licensed use limited to: Odisha University of Technology and Research. Downloaded on October 24,2024 at 08:10:04 UTC from IEEE Xplore. Restrictions apply.
 158 IEEE SIGNAL PROCESSING LETTERS, VOL. 27, 2019

Fig. 3. Overall flow of the proposed method. Given labeled HR images, domain discriminator in generative face augmentation (GFA) encourages attention-guided
generator to transfer them into LR-like images at image-level. Spatial resolution adaptation (SRA) ensures that LR-like images are indistinguishable from real LR
images at feature-level. Note that the solid line is the path of the labeled HR images, while the dotted line is the path of the unlabeled LR images.

image, i.e., foreground. Given an HR image, we first feed it to the where, LiC and LiD refer to the loss of the i-th sample from
generator and attention module. We can generate the foreground class and domain predictions, respectively. The parameter λ is
area via a Hadamard product [41] between domain-transferred the most critical factor in which a negative sign of λ leads to an
image and attention map. After that, we obtain the background adversarial relationship between F and D in terms of loss. As a
area from the counterpart of the attention map, and add them to result, the parameters of F converge at a compromise point that
the image with the foreground transferred. As a result, we can is discriminative and also satisfies domain invariance through
obtain a domain transferred image as: the process of minimizing network loss.

hr − lr = GenHR→LR (hr)  mAtt +hr  (1 − mAtt ), (1) IV. EXPERIMENTAL RESULTS

A. Experimental Setup
where hr is a image from source HR domain and mAtt refers
to attention map derived from attention module AttHR−>LR To construct GFA, we build on the ‘pix2pix’ framework in
containing the value [0, 1] per pixel. which ‘generator’ consists of convolutions, residual blocks, and
deconvolutions, and ‘discriminator’ is a two-class classification
network with several convolutions. We apply several augmenta-
B. Spatial Resolution Adaptation (SRA)
tion schemes (e.g., flip, rotation, and crop) [42], [43] to prevent
Although we generate image-level domain-transferred im- over-fitting and DA failure due to the lack of source domain
ages, there still exists a difference between source HR domain images. Further, when pose variation is severe in the target
and target LR domain at feature-level. Therefore, we introduce domain, we optionally perform 3D image synthesis [24]. For
an SRA network that predicts class labels and domain labels SRA, we adopt pre-trained VGG-Face [1] as a feature extractor,
(i.e., HR/LR) as can be seen in Fig. 3. SRA aims to learn the attach shallow networks for classifier and discriminator, and
unified feature distribution that is discriminative in both HR and finally fine-tune them. To prevent over-fitting on relatively small
LR domains without LR domain labels. For the unified feature datasets, we freeze the weights from the Convl layer to FC6 layer
distribution, we update the parameters of F and C, θF and θC , to of the pre-trained VGG-Face and fine-tune the FC7 layer. For
minimize the label prediction loss in HR domain. Here, F indi- more details of SRA, please refer to the project page1 of our
cates feature extractor and C refers to the class-label classifier. prior work [24]
Concurrently with the discriminant learning, we should align We demonstrate the effectiveness of the proposed method on
the distribution between HR domain and LR domain. In order public datasets, EK-LFH [24], SCface [27], and YouTubeFaces
to obtain domain-invariant features, SRA struggles to find a θF (YTF) [28]. Details of each dataset are presented in Table I.
that maximizes the domain prediction loss, while simultaneously Considering real-world scenarios, we follow the subject-disjoint
searching for parameters of domain discriminator D (θD ) to protocol in LRFR [10], [29], [44], which prevents the use of a
minimize domain prediction loss as follows: test image of subjects that appear in the training phase.
 
L= LiC + LiD when update θD , B. Evaluation on EK-LFH
i∈HR i∈HR ∪ LR
  Following the subject-disjoint protocol [10], [29], [44] in EK-
L= LiC − λ LiD when update θF and θC , LFH [24], we use 20 randomly selected subjects for training
i∈HR i∈HR ∪ LR

(2) 1 https://github.com/csehong/SSPP-DAN

Authorized licensed use limited to: Odisha University of Technology and Research. Downloaded on October 24,2024 at 08:10:04 UTC from IEEE Xplore. Restrictions apply.
 HONG AND RYU: UNSUPERVISED FACE DOMAIN TRANSFER FOR LOW-RESOLUTION FACE RECOGNITION 159

TABLE I TABLE III

DATASET SPECIFICATION. TWO NUMBERS OF THE TARGET DOMAIN IN THE ABLATION ANALYSIS ON SCface
SUBJECT COLUMN INDICATE THE NUMBER OF SUBJECTS USED FOR
TRAINING AND TEST, RESPECTIVELY

TABLE II
ABLATION ANALYSIS ON EK-LFH. SUBSCRIPTS ‘L’ AND ‘U’ INDICATE TABLE IV
LABELED AND UNLABELED IMAGES AND ‘POSE’ REFERS TO SYNTHESIZED COMPARISON BETWEEN MODELS ON SCFACE
IMAGES WITH VARIOUS POSES

label information in the target LR domain, the proposed method

outperforms the supervised LRFR methods. Finally, we can see
that the proposed method including image-level domain transfer
and the other 10 subjects for test. Pose variation is severe in
is superior to our previous work based on unsupervised DA [24].
the target domain on EK-LFH as can be seen in Table I, thus,
we perform 3D image synthesis [24] during image-level domain
transfer. Motivated by the experimental settings in conventional D. Evaluation on YouTubeFaces
unsupervised DA [24], [45], we perform experiments on three
We perform the evaluation on a large-scale YTF dataset [28]
protocols as shown in Table II. First, ‘Source Only’ protocol only
to demonstrate the generalization of the proposed method. YTF
uses labeled samples in the source HR domain, which revealed
dataset consists of 3,425 videos of 1,595 subjects, including
the theoretical lower bound on performance as 20.97%. We also
a total of 620 k images. We resize the first frame of each
report the results of adding our synthesized images (e.g., HRpose
video 224 × 224 for the HR domain while downsampling the
or HR-LR) to the training set as variants of conventional ‘Source
remaining frame to 16 × 16 for the LR domain. Among the
Only.’ Second, as an upper performance bound, the model from
downsampled LR domain samples, we divide them in half
‘Labeled Target’ protocol is trained by the labeled source HR
set and use each set for the train and test data respectively.
images as well as labeled target LR images.
We compare the proposed adaptation method with the ‘Source
Finally, our main approaches, ‘Proposed Adaptation,’ use
Only’ baseline under the same experimental settings of EK-LFH
labeled source images with unlabeled target images. We observe
and SCface datasets. We confirm that the ‘Proposed Adap-
that the proposed method with synthesized images improves
tation’ achieves 19.2%, which is much higher than ‘Source
accuracy compared to ‘Source Only,’ even though the labels of
Only’ with 8.0% accuracy. This result supports the general-
the target domain are not used. The fifth and sixth rows validate
ization performance of the proposed method in a large-scale
the importance of synthesized images with various poses when
dataset.
applying unsupervised DA. Also, we confirm that the proposed
image-level domain transfer can improve performance even
further. Overall results show that GFA (image-level) and SRA V. CONCLUSION
(feature-level) work complementarily in solving the LRFR task.
A combined LRFR model with generative face augmentation
(GFA) and spatial resolution adaptation (SRA) is proposed to
C. Evaluation on SCface reduce the domain differences at feature-level as well as image-
Following the existing protocol in SCface [9], [10], [29], level. In contrast to conventional methods, the proposed method
we split the training set and test set into 50 and 80 subjects does not use any labels on LR probe images to account for
separately. From Table III, we confirm a considerable perfor- real-world face recognition. For image-level domain transfer, we
mance improvement by using ‘Proposed Adaptation’ even if no first apply the proposed attention-guided GFA to the source HR
label information of the target domain is used. It is apparent images. Image-level domain transferred images are then trained
that the experimental results on SCface are consistent with that in SRA to generate domain-invariant and discriminant feature
of the EK-LFH dataset. Finally, we compare our method with distributions. Our unsupervised method of jointly training GFA
the state-of-the-art supervised methods that use labels in the and SRA outperforms the state-of-the-art supervised methods
target domain, as can be seen in Table IV. Despite not using any on the public dataset.

Authorized licensed use limited to: Odisha University of Technology and Research. Downloaded on October 24,2024 at 08:10:04 UTC from IEEE Xplore. Restrictions apply.
 160 IEEE SIGNAL PROCESSING LETTERS, VOL. 27, 2019

REFERENCES [23] Q. Qiu, J. Ni, and R. Chellappa, “Dictionary-based domain adapta-

tion methods for the re-identification of faces,” in Proc. Person Re-
[1] O. M. Parkhi et al., “Deep face recognition,” in Proc. British Mach. Vision Identification, 2014, pp. 269–285.
Conf., vol. 1, no. 3, 2015, p. 6. [24] S. Hong, W. Im, J. Ryu, and H. S. Yang, “Sspp-dan: Deep domain
[2] S. Hong, W. Im, J. Park, and H. Yang, “Deep cnn-based person identi- adaptation network for face recognition with single sample per person,” in
fication using facial and clothing features,” in Proc. Summer Conf. Inst. Proc. IEEE Int. Conf. Image Process., 2017, pp. 825–829.
Electron. Inf. Engineers, 2016, pp. 2204–2207. [25] M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative adver-
[3] Y. Zhong, J. Chen, and B. Huang, “Toward end-to-end face recognition sarial networks,” in Proc. Int. Conf. Mach. Learn., 2017, pp. 214–223.
through alignment learning,” IEEE Signal Process. Lett., vol. 24, no. 8, [26] I. Goodfellow, “Nips 2016 tutorial: Generative adversarial networks,”
pp. 1213–1217, Aug. 2017. 2016, arXiv:1701.00160.
[4] S. Gao, Y. Zhang, K. Jia, J. Lu, and Y. Zhang, “Single sample face [27] M. Grgic, K. Delac, and S. Grgic, “Scface–surveillance cameras face
recognition via learning deep supervised autoencoders,” IEEE Trans. Inf. database,” Multimedia Tools Appl., vol. 51, no. 3, pp. 863–879, 2011.
Forensics Secur., vol. 10, no. 10, pp. 2108–2118, Oct. 2015. [28] L. Wolf, T. Hassner, and I. Maoz, “Face recognition in unconstrained
[5] S. Gao, K. Jia, L. Zhuang, and Y. Ma, “Neither global nor local: videos with matched background similarity,” in Proc. Comput. Vision
Regularized patch-based representation for single sample per person Pattern Recognit., 2011, pp. 529–534.
face recognition,” Int. J. Comput. Vision, vol. 111, no. 3, pp. 365–383, [29] S. P. Mudunuri, S. Venkataramanan, and S. Biswas, “Dictionary alignment
2015. with re-ranking for low-resolution nir-vis face recognition,” IEEE Trans.
[6] O. Abdollahi Aghdam, B. Bozorgtabar, H. Kemal Ekenel, and J.-P. Thiran, Inf. Forensics Secur., vol. 14, no. 4, pp. 886–896, Apr. 2019.
“Exploring factors for improving low resolution face recognition,” in Proc. [30] Y. Chu, T. Ahmad, G. Bebis, and L. Zhao, “Low-resolution face recognition
Comput. Vision Pattern Recognit. Workshops, 2019. with single sample per person,” Signal Process., vol. 141, pp. 144–157,
[7] Z. Cheng, X. Zhu, and S. Gong, “Surveillance face recognition challenge,” 2017.
2018, arXiv:1804.09691. [31] M. Haghighat and M. Abdel-Mottaleb, “Low resolution face recognition
[8] P. Li, L. Prieto, D. Mery, and P. Flynn, “Face recognition in low quality in surveillance systems using discriminant correlation analysis,” in Proc.
images: A survey,” 2018, arXiv:1805.11519. IEEE Int. Conf. Autom. Face Gesture Recognit., 2017, pp. 912–917.
[9] P. Li, L. Prieto, D. Mery, and P. J. Flynn, “On low-resolution face recog- [32] X. Yu, B. Fernando, R. Hartley, and F. Porikli, “Super-resolving very low-
nition in the wild: Comparisons and new techniques,” IEEE Trans. Inf. resolution face images with supplementary attributes,” in Proc. Comput.
Forensics Secur., vol. 14, no. 8, pp. 2000–2012, Aug. 2019. Vision Pattern Recognit., 2018, pp. 908–917.
[10] S. Ge, S. Zhao, C. Li, and J. Li, “Low-resolution face recognition in the [33] X. Xie, Z. Cao, Y. Xiao, M. Zhu, and H. Lu, “Blurred image recognition
wild via selective knowledge distillation,” IEEE Trans. Image Process., using domain adaptation,” in Proc. IEEE Int. Conf. Image Process., 2015,
vol. 28, no. 4, pp. 2051–2062, Apr. 2019. pp. 532–536.
[11] Z. Wang, S. Chang, Y. Yang, D. Liu, and T. S. Huang, “Studying very [34] S. Banerjee and S. Das, “Domain adaptation with soft-margin multiple
low resolution recognition using deep networks,” in Proc. Comput. Vision feature-kernel learning beats deep learning for surveillance face recogni-
Pattern Recognit., 2016, pp. 4792–4800. tion,” 2016, arXiv:1610.01374.
[12] B. F. Klare and A. K. Jain, “Heterogeneous face recognition using kernel [35] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation
prototype similarities,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 35, with conditional adversarial networks,” in Proc. Comput. Vision Pattern
no. 6, pp. 1410–1422, Jun. 2013. Recognit., 2017, pp. 1125–1134.
[13] S. Liao, D. Yi, Z. Lei, R. Qin, and S. Z. Li, “Heterogeneous face recogni- [36] I. Goodfellow et al., “Generative adversarial nets,” in Proc. Neural Inf.
tion from local structures of normalized appearance,” in Proc. Int. Conf. Process. Syst., 2014, pp. 2672–2680.
Biometrics, 2009, pp. 209–218. [37] J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image
[14] B. Li, H. Chang, S. Shan, and X. Chen, “Low-resolution face recognition translation using cycle-consistent adversarial networks,” in Proc. Int. Conf.
via coupled locality preserving mappings,” IEEE Signal Process. Lett., Comput. Vision, 2017, pp. 2223–2232.
vol. 17, no. 1, pp. 20–23, Jan. 2010. [38] Y. A. Mejjati, C. Richardt, J. Tompkin, D. Cosker, and K. I. Kim, “Un-
[15] P. Li, J. Brogan, and P. J. Flynn, “Toward facial re-identification: Experi- supervised attention-guided image-to-image translation,” in Proc. Neural
ments with data from an operational surveillance camera plant,” in Proc. Inf. Process. Syst., 2018, pp. 3693–3703.
IEEE Int. Conf. Biometrics Theory, Appl. Syst., 2016, pp. 1–8. [39] S. Woo, J. Park, J.-Y. Lee, and I. So Kweon, “Cbam: Convolutional block
[16] P. Li, M. L. Prieto, P. J. Flynn, and D. Mery, “Learning face similarity for attention module,” in Proc. Eur. Conf. Comput. Vision, 2018, pp. 3–19.
re-identification from real surveillance video: A deep metric solution,” in [40] F. Wang et al., “Residual attention network for image classification,” in
Proc. IEEE Int. Joint Conf. Biometrics, 2017, pp. 243–252. Proc. Comput. Vision Pattern Recognit., 2017, pp. 3156–3164.
[17] Z. Lu, X. Jiang, and A. Kot, “Deep coupled resnet for low-resolution [41] R. A. Horn, “The hadamard product,” in Proc. Symp. Appl. Math, vol. 40,
face recognition,” IEEE Signal Process. Lett., vol. 25, no. 4, pp. 526–530, 1990, pp. 87–169.
Apr. 2018. [42] W. Im, S. Hong, S.-E. Yoon, and H. S. Yang, “Scale-varying triplet ranking
[18] M. Jian and K.-M. Lam, “Simultaneous hallucination and recognition with classification loss for facial age estimation,” in Proc. Asian Conf.
of low-resolution faces based on singular value decomposition,” ACM Comput. Vision. Springer, 2018, pp. 247–259.
Trans. Circuits Syst. Video Technol., vol. 25, no. 11, pp. 1761–1772, [43] S. Datta, G. Sharma, and C. Jawahar, “Unsupervised learning of face
2015. representations,” in Proc. IEEE Int. Conf. Autom. Face Gesture Recognit.,
[19] M.-C. Yang, C.-P. Wei, Y.-R. Yeh, and Y.-C. F. Wang, “Recognition at a 2018, pp. 135–142.
long distance: Very low resolution face recognition and hallucination,” in [44] Z. Wang, D. Liu, J. Yang, W. Han, and T. Huang, “Deep networks for
Proc. IEEE Int. Conf. Biometrics, 2015, pp. 237–242. image super-resolution with sparse prior,” in Proc. Int. Conf. Comput.
[20] S. Kolouri and G. K. Rohde, “Transport-based single frame super reso- Vision, 2015, pp. 370–378.
lution of very low resolution face images,” in Proc. of Comput. Vision [45] Y. Ganin and V. Lempitsky, “Unsupervised domain adaptation by back-
Pattern Recognit., 2015, pp. 4876–4884. propagation,” in Proc. of Int. Conf. Mach. Learn., 2015, pp. 1180–1189.
[21] C.-X. Ren, D.-Q. Dai, and H. Yan, “Coupled kernel embedding for low- [46] J. Yang, J. Wright, T. S. Huang, and Y. Ma, “Image super-resolution
resolution face image recognition,” IEEE Trans. Image Process., vol. 21, via sparse representation,” IEEE Trans. Image Process., vol. 19, no. 11,
no. 8, pp. 3770–3783, Aug. 2012. pp. 2861–2873, Nov. 2010.
[22] S. Shekhar, V. M. Patel, and R. Chellappa, “Synthesis-based recognition [47] Z. Wang, W. Yang, and X. Ben, “Low-resolution degradation face recog-
of low resolution faces,” in Proc. IEEE Int. Joint Conf. Biometrics, 2011, nition over long distance based on cca,” Neural Comput. Appl., vol. 26,
pp. 1–6. no. 7, pp. 1645–1652, 2015.

Authorized licensed use limited to: Odisha University of Technology and Research. Downloaded on October 24,2024 at 08:10:04 UTC from IEEE Xplore. Restrictions apply.