Attention-based Gender Recognition on Masked Faces

Vincenzo Carletti, Antonio Greco, Alessia Saggese and Mario Vento

Dept. of Information Eng., Electrical Eng. and Applied Mathematics (DIEM), University of Salerno, Italy

Keywords:

Gender Recognition, Attention Mechanism, Masked Faces.

Abstract:

Gender recognition from face images can be proﬁtably used in several vertical markets, such as targeted

advertising and cognitive robotics. However, in the last years, due to the COVID-19 pandemic, the unreliability

of such systems when dealing with faces covered by a mask has emerged. In this paper, we propose a novel

architecture based on attention layers and trained with a domain speciﬁc data augmentation technique for

reliable gender recognition of masked faces. The proposed method has been experimentally evaluated on a

huge dataset, namely VGGFace2-M, a masked version of the well known VGGFace2 dataset, and the achieved

results conﬁrm an improvement of around 4% with respect to traditional gender recognition algorithms, while

preserving the performance on unmasked faces.

1 INTRODUCTION

Face analysis, such as gender (Ng et al., 2015) and

ethnicity recognition (Greco et al., 2020a) or age esti-

mation (Carletti et al., 2020), attracted a growing in-

terest of the scientiﬁc community in the last decade.

Indeed, the above tasks can be considered as the base

for a huge amount of different applications, such as

targeted advertising (Greco et al., 2020b), forensic

search for video surveillance or social robots (Foggia

et al., 2019). Among the different face analysis tasks,

in this paper we will focus on the gender recognition

problem.

In the last years plenty of methods have been

proposed for gender recognition from face images,

which adopt color, shape and texture features (Zhang

et al., 2016) (Azzopardi et al., 2016a) (Azzopardi

et al., 2017) (Carletti et al., 2020), trainable fea-

tures (Azzopardi et al., 2016b) (Simanjuntak and Az-

zopardi, 2019) (Azzopardi et al., 2018a), a combina-

tion of handcrafted and trainable features (Azzopardi

et al., 2018b), single convolutional neural networks

(CNNs) (Levi and Hassner, 2015) (Antipov et al.,

2017) (Greco et al., 2020), multi-task CNNs (Ran-

jan et al., 2017) (Dehghan et al., 2017) (Gurnani

et al., 2019) or ensemble of CNNs (Aﬁﬁ and Ab-

delhamed, 2019) (Antipov et al., 2016). These ap-

proaches achieve over 95% of accuracy on facial im-

ages collected in controlled environments, but their

performance decrease in presence of strong varia-

tions (pose, blur, noise, partial occlusions and so on).

To further investigate the effect of image corruptions

over gender recognition algorithms, a framework has

been recently proposed and made publicly available

(Carletti et al., 2020).

The spread of the COVID-19 pandemic has made

the gender recognition problem even more difﬁcult,

since the masks occlude part of the faces. A deep

analysis, involving seven different deep architecture

devised for gender recognition, has been conducted

in (Greco et al., 2021), where it was demonstrated

that the performance drop for this speciﬁc problem

is about 10%. Even if this is not dramatic, like in

the case of emotion recognition, for which the accu-

racy drop is instead about 50%, some effort is still

required, aiming at improving the performance of

gender recognition algorithms in presence of masked

faces, but mostly important without degrading the

performance in case of unmasked faces.

The problem of occlusions, also including scarves

or glasses, is among the main challenges of face anal-

ysis tasks, and in recent years several solutions have

been proposed for solving this problem, especially in

face recognition tasks. One of the ﬁrst approaches for

dealing with occlusions has been proposed in (Min

et al., 2011): the face ROI is ﬁrstly split into two

horizontal half and, for each part, the presence of

scarf/sunglasses is detected by Gabor wavelets, let-

ting the model processing only the non-occluded fa-

cial regions with a SVM classiﬁer. In 2019 and in

2020, an increasing number of face detection and

recognition and expression analysis algorithm have

672

Carletti, V., Greco, A., Saggese, A. and Vento, M.

Attention-based Gender Recognition on Masked Faces.

DOI: 10.5220/0010978700003124

In Proceedings of the 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2022) - Volume 5: VISAPP, pages

672-678

ISBN: 978-989-758-555-5; ISSN: 2184-4321

been proposed for dealing with occlusions, due to the

COVID-19 pandemic. A common trend has been to

introduce in the architectures visual attention mecha-

nisms (Li et al., 2019)(Xu et al., 2020)(Yuan, 2020).

The attention tries to emulate a cognitive process

adopted by the human brain when it focuses only on

some areas of the image most promising for the ex-

traction of the salient features concerning the problem

to be solved. To emulate this behaviour, the attention

mechanism associates different weights to the differ-

ent features extracted by the network.

Another commonly adopted strategy has been the

use of augmentation techniques for improving age

and gender recognition algorithms in presence of oc-

cluded faces (Hsu et al., 2021). The authors introduce

three occlusion techniques, namely blackout, random

brightness and blur, designed to simulate different

challenges that could be experienced in real-world ap-

plications. Anyway, even if interesting, these policies

are not domain dependent and, thus, can not really

solve the issue related to the presence of the mask

covering the face. The idea could be instead to aug-

ment the training set with some domain speciﬁc poli-

cies, such as faces covered by a mask, obtained by

adding synthetic occlusions to the image.

Starting from the above considerations, in this pa-

per we propose a reliable gender recognition algo-

rithm, based on the attention mechanism and trained

with a domain speciﬁc data augmentation technique,

able to achieve a remarkable accuracy both in pres-

ence and in absence of masks occluding the faces.

According to our knowledge, this is the ﬁrst time that

domain speciﬁc techniques data augmentation and at-

tention mechanisms are used together to improve per-

formance over masked faces of gender recognition al-

gorithms.

2 THE PROPOSED APPROACH

The method is based on the standard processing

pipeline for these types of systems: face detection and

gender recognition. The architecture of the proposed

approach is depicted in Figure 1.

In the face detection step we localize the face in

the whole image and crop the region that must be

given as input to the classiﬁer. The most popular face

detectors recently demonstrated a great accuracy (Li

et al., 2021), but they are not so effective on faces cov-

ered by a mask, since the common datasets adopted

for face detection do not include masked faces. There-

fore, for our method we trained our own face detector

based on MobileNetv2-SSD (Sandler et al., 2018), by

using all the images available in the training sets of

WiderFace (Yang et al., 2016) and MAFA (Ge et al.,

2017); the evaluation of the accuracy of this module

is out of the scope of this paper, but it has been able

to locate and crop all the masked and unmasked faces

in the dataset we used to train, validate and test our

method.

In the gender recognition step we apply our CNN

on the face image, resized to 224×224. The proposed

CNN is a custom version of ResNet50 (He et al.,

2016), in which we have modiﬁed the Residual Unit

(RU) to add an attention module. In particular, we

adopt the visual attention mechanism (Liu and Mi-

lanova, 2018) to emulate the way in which the human

brain works: humans are able to dynamically locate

the region of interest and analyze the scene by selec-

tively processing subsets of the visual input. In the

visual attention mechanisms, the main idea is to as-

sociate different weights to the different features ex-

tracted by the network, with the aim to focus not on

the whole image but instead only on some areas of

the image, namely the most promising ones. For our

problem at hands, the reason why we propose to in-

troduce attention layer is to use the whole face when

available, while only the top part of the face when it

is covered by a mask.

To this aim, we followed the method proposed in

(Hu et al., 2018), namely we weigh the output feature

block of the RU with weights computed by an addi-

tional attention module (see Figure 1). There are dif-

ferent visual attention mechanisms, but we selected

the ones that use or combine channel and spatial at-

tention, that are suited for our purposes. In partic-

ular, we chose three of them, widely adopted in the

scientiﬁc literature, namely the Convolutional Block

Attention Module (CBAM) (Woo et al., 2018), the

spatial and channel Squeeze and Excitation (scSE)

(Roy et al., 2018) and the Efﬁcient Channel Attention

(ECA) (Wang et al., 2020).

CBAM infers, given an initial feature map, the at-

tention weights along the channel and spatial dimen-

sions separately. The computed attention maps are

then multiplied with the input feature map to reﬁne

the features. scSE still uses both channel and spatial

dimensions, but concurrently; indeed, scSE computes

spatial and channel-wise attention maps by concur-

rently recalibrating the input spatially and channel-

wise. This module is an extended version of the

well known SE attention module, which is the base

of SENet architecture (Hu et al., 2018). It is impor-

tant to highlight that both CBAM and scSE compute

spatial attention using a 2D convolution of kernel size

K × K. Differently, ECA investigates a 1D convolu-

tion with adaptive kernel size to replace the fully con-

nected layers in the channel attention module.

Attention-based Gender Recognition on Masked Faces

673

Figure 1: Overview of the proposed method. Face detection is performed with a custom detector based on MobileNet-SSD

and trained with masked and unmasked faces. Gender recognition is carried out with a modiﬁed version of ResNet50, in

which the Residual Unit is extended with an attention module; this change allows to dynamically focus on the visible portion

of the face.

3 EXPERIMENTAL RESULTS

In this Section we will detail the dataset used in our

experiments (see Subsection 3.1), the training proce-

dure (see Subsection 3.2) and ﬁnally the achieved re-

sults (see Subsection 3.3) and a visual explanation of

the experimental ﬁndings (see Subsection 3.4).

3.1 Dataset

We conducted our experimentation over the

VGGFace2-M dataset, that we recently proposed

in (Greco et al., 2021). As well as the original

VGGFace2 (Cao et al., 2018) dataset, it contains

over 3 millions of face images belonging to 9, 131

different subjects, annotated with gender.

Basically, it includes the same face images of

the original dataset, but covered with a synthetically

added mask. It has been produced by using the ap-

proach proposed in (yuan Wang et al., 2020): the dlib

face detector (Kazemi and Sullivan, 2014) has been

applied and 68 facial landmarks have been identiﬁed

on each face. Thus, the following four points have

been used to determine the position and the shape

of the mask on the face, namely the nose point, the

chin bottom point, the chin right point and the chin

left point. The left part of the mask is built by using

Figure 2: Images from the VGGFace2-M dataset.

the chin left point, the chin bottom point and the nose

point. The right part is obtained symmetrically with

the chin right point. The two parts are then merged, by

obtaining a mask whose height is equal to the distance

between the nose point and the bottom chin point and

whose width is equal to the sum of the widths of the

two parts. The mask is ﬁnally rotated of the angle be-

tween the nose point and the extremity of one of the

eyes. Some example of masked faces belonging to the

VGGFace2-M dataset are reported in Figure 2.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

674

3.2 Training Procedure

The proposed architecture has been trained through a

SGD optimizer (Liu et al., 2020). The weights have

been pre-trained over the ImageNet dataset and a ﬁne

tuning of all the layers has been performed. The net-

work has been trained for 70 epochs with batches of

128 samples and a learning rate starting from 0.005

with a decay of factor equal to 5 every 20 epochs. The

adopted loss function is the categorical cross-entropy,

with a regularization weight decay of 0.005.

Different types of general purpose data augmen-

tation strategies commonly adopted for face analysis

tasks are applied: random rotation (±10

◦

), shear (up

to 10%), cropping (up to 5% variation over the origi-

nal face region), horizontal ﬂipping (with 50% proba-

bility), change of brightness (up to ±20% of the avail-

able range) and contrast(up to ±50% of the maximum

value). The chosen ranges have been experimentally

chosen in order to reproduce conditions that are likely

to occur in real scenarios. Also, a domain speciﬁc

data augmentation technique has been performed: in-

deed, masked faces from the VGGFace2-M dataset

have been included in the dataset for training.

3.3 Overall Results

The obtained results are summarized in Table 1,

where we report the results in terms of accuracy

on the test set of both the datasets VGGFace2 and

VGGFace2-M. The drop over masked faces with re-

spect to unmasked face is also in the table (Drop), to-

gether with the average accuracy value (Avg).

Also, as we can see from column Training set, two

different policies have been considered for training:

• VGGFace2-M: only the masked version of VG-

GFace2 dataset, namely VGGFace2-M, without

any unmasked face, has been used for training. In

this case, we suppose that any subject is wearing

a mask. Anyway, we also evaluated the perfor-

mance over unmasked faces.

• VGGFace2-UM: the union of the two datasets,

namely VGGFace2 and VGGFace2-M, have been

used for training.

We also include in the table results achieved by

the baseline solution, namely the model with the best

performance over both VGGFace2 and VGGFace2-

M, as reported in (Greco et al., 2021). It is a VGG-

16 CNN trained for gender recognition over the VG-

GFace2 dataset. We can note that this result is among

the best ones on VGGFace2 (97.60%), but also, as ex-

pected, the worst one on masked faces (92.99%), with

a performance drop of 4.61%.

In average, the best result is obtained when

ResNet50+CBAM model is trained on VGGFace2-

UM, achieving the best average accuracy of 97.23%

and the best performance over unmasked faces of

97.61%, with a drop of 0.76% (vs 4.61% of the base-

line solution). It is important to highlight that the

accuracy achieved over the unmasked VGGFace2 is

also higher than the VGG baseline network trained

over the unmasked training set. Even if this ar-

chitecture does not achieve the best result over

masked faces, it is very close to it, being 96.85% vs

97.02% the accuracy achieved by the same architec-

ture (ResNet50+CBAM) trained with VGGFace2-M.

Also, it is important to note that all the architec-

tures based on attention layers achieve better results

than the baseline solution. This achievement further

conﬁrms that the attention mechanism allows to si-

multaneously and reliably manage both masked and

unmasked faces, reaching state of the art accuracy

when dealing with unmasked faces but reporting only

a very slight drop in the performance (lower than 1%)

also over masked faces. Therefore, it can be consid-

ered deﬁnitively a viable solution for the problem at

hand.

Figure 3: Examples of the proposed system in action. M

and F in the image indicate Male and Female, respectively.

For each face, the presence of the mask is represented with

a green bounding box, while the absence is indicated with a

red rectangle.

3.4 Explainability of the Results

In this section we aim to explain the results obtained

by the proposed network through the visualization of

the discriminative features learned by the CNN. To

this purpose, we computed the class activation maps

(Zhou et al., 2016) to identify the region of the face

mostly used by the CNNs for recognizing the gen-

der. The class activation map is a heat map computed

for every pixel of the image, in which the ”hot” pix-

els, coloured with red gradations, represent regions

of the image mostly used by the CNN to extract dis-

criminative features for the classiﬁcation. We used

Grad-CAM (Selvaraju et al., 2017) to compute the

Attention-based Gender Recognition on Masked Faces

675

Table 1: Accuracy achieved over both VGGFace2 and VGGFace2-M by varying the speciﬁc model considered and the training

set. In bold we report the best result, namely the best accuracy over VGGFace2 and VGGFace2-M, the minimum drop and

the highest average accuracy.

Model Training set

Accuracy (%)

VGGFace2 VGGFace2-M Drop Avg.

VGG VGGFace2 97.60 92.99 4.61 95.30

VGG VGGFace2-M 95.72 96.37 -0.66 96.05

ResNet50+ECA VGGFace2-M 96.58 96.87 -0.29 96.73

ResNet50+CBAM VGGFace2-M 96.81 97.02 -0.21 96.92

ResNet50+scSE VGGFace2-M 96.76 96.96 -0.20 96.86

ResNet50+ECA VGGFace2-UM 97.54 96.77 0.78 97.15

ResNet50+CBAM VGGFace2-UM 97.61 96.85 0.76 97.23

ResNet50+scSE VGGFace2-UM 97.54 96.78 0.76 97.16

(a) Reference (b) VGG-U (c) VGG-M (d) CBAM-M (e) CBAM-UM

Figure 4: Class activation maps computed on (a) male and female unmasked and masked reference faces for (b) VGG16

trained on VGGFace2, (c) VGG16 trained on VGGFace2-M, (d) ResNet50+CBAM trained on VGGFace2-M and (e)

ResNet50+CBAM trained on VGGFace2-UM.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

676

class activation maps of the two versions of VGG

and ResNet50+CBAM and reported some examples

in Figure 4.

It is evident that the two versions of VGG do

not use the whole part of the face visible in pres-

ence of mask; the version trained with VGGFace2-M

not only suffer of the above mentioned problem, but

also uses very small parts of the face even when it is

completely visible. On the other hand, the proposed

ResNet50+CBAM is able to focus the attention on the

whole region of the face that is visible; in presence of

mask, it uses the region of the eyes and the upper part

of the nose, while on unmasked faces the CNN takes

advantage of the features of the whole face to perform

the classiﬁcation. It is particularly evident on the ver-

sion trained with VGGFace2-UM, while the other fo-

cuses more on the upper part of the face, being trained

only with masked faces.

4 CONCLUSIONS

State of the art algorithms for gender recognition from

face images, even if very reliable, suffer a perfor-

mance drop in presence of occlusions covering the

face. This has a great impact especially during the

COVID-19 pandemic, when wearing a mask became

mandatory in several countries all around the world.

Starting from this assumption, in this paper we have

presented a novel algorithm exploiting the attention

mechanism and a domain speciﬁc data augmentation

policy able to achieve on average a 4% improve-

ment with respect to traditional state of the art gen-

der recognition algorithm, while preserving the per-

formance over unmasked faces. It is due to the capa-

bility of the network, demonstrated through a visual

investigation with class activation maps, to adaptively

select the discriminative region of interests, namely

the visible parts of the face.

Future works include the extension of this work to

real masked faces and to other face analysis tasks that

still suffer from the presence of the mask, such as age

estimation or ethnicity recognition.

REFERENCES

Aﬁﬁ, M. and Abdelhamed, A. (2019). Aﬁf4: deep gender

classiﬁcation based on adaboost-based fusion of iso-

lated facial features and foggy faces. Journal of Visual

Communication and Image Representation, 62:77–86.

Antipov, G., Baccouche, M., Berrani, S.-A., and Dugelay,

J.-L. (2017). Effective training of convolutional neu-

ral networks for face-based gender and age prediction.

Pattern Recognition, 72:15–26.

Antipov, G., Berrani, S.-A., and Dugelay, J.-L. (2016). Min-

imalistic cnn-based ensemble model for gender pre-

diction from face images. Pattern Recognition Letters,

70:59–65.

Azzopardi, G., Foggia, P., Greco, A., Saggese, A., and

Vento, M. (2018a). Gender recognition from face im-

ages using trainable shape and color features. In 2018

24th International Conference on Pattern Recognition

(ICPR), pages 1983–1988. IEEE.

Azzopardi, G., Greco, A., Saggese, A., and Vento, M.

(2017). Fast gender recognition in videos using a

novel descriptor based on the gradient magnitudes of

facial landmarks. In 2017 14th IEEE International

Conference on Advanced Video and Signal Based

Surveillance (AVSS), pages 1–6. IEEE.

Azzopardi, G., Greco, A., Saggese, A., and Vento, M.

(2018b). Fusion of domain-speciﬁc and trainable fea-

tures for gender recognition from face images. IEEE

Access, 6:24171–24183.

Azzopardi, G., Greco, A., and Vento, M. (2016a). Gen-

der recognition from face images using a fusion of

svm classiﬁers. In International Conference on Image

Analysis and Recognition, pages 533–538. Springer.

Azzopardi, G., Greco, A., and Vento, M. (2016b). Gen-

der recognition from face images with trainable cos-

ﬁre ﬁlters. In IEEE International Conference on Ad-

vanced Video and Signal Based Surveillance (AVSS),

pages 235–241. IEEE.

Cao, Q., Shen, L., Xie, W., Parkhi, O. M., and Zisserman,

A. (2018). Vggface2: A dataset for recognising faces

across pose and age. In 2018 13th IEEE International

Conference on Automatic Face Gesture Recognition

(FG 2018), pages 67–74.

Carletti, V., Greco, A., Percannella, G., and Vento, M.

(2020). Age from faces in the deep learning revolu-

tion. IEEE Transactions on Pattern Analysis and Ma-

chine Intelligence, 42(9):2113–2132.

Carletti, V., Greco, A., Saggese, A., and Vento, M. (2020).

An effective real time gender recognition system for

smart cameras. J. Ambient Intell. Humaniz. Comput.,

11(6):2407–2419.

Dehghan, A., Ortiz, E. G., Shu, G., and Masood, S. Z.

(2017). Dager: Deep age, gender and emotion recog-

nition using convolutional neural network. arXiv

preprint arXiv:1702.04280.

Foggia, P., Greco, A., Percannella, G., Vento, M., and Vig-

ilante, V. (2019). A system for gender recognition on

mobile robots. In Proceedings of the 2nd Interna-

tional Conference on Applications of Intelligent Sys-

tems, pages 1–6.

Ge, S., Li, J., Ye, Q., and Luo, Z. (2017). Detecting masked

faces in the wild with lle-cnns. In Proceedings of

the IEEE Conference on Computer Vision and Pattern

Recognition, pages 2682–2690.

Greco, A., Percannella, G., Vento, M., and Vigilante,

V. (2020a). Benchmarking deep network architec-

tures for ethnicity recognition using a new large face

dataset. Machine Vision and Applications.

Attention-based Gender Recognition on Masked Faces

677

Greco, A., Saggese, A., and Vento, M. (2020b). Digi-

tal signage by real-time gender recognition from face

images. In 2020 IEEE International Workshop on

Metrology for Industry 4.0 & IoT, pages 309–313.

Greco, A., Saggese, A., Vento, M., and Vigilante, V. (2020).

A convolutional neural network for gender recogni-

tion optimizing the accuracy/speed tradeoff. IEEE Ac-

cess, 8:130771–130781.

Greco, A., Saggese, A., Vento, M., and Vigilante, V. (2021).

Performance assessment of face analysis algorithms

with occluded faces. In International Conference on

Pattern Recognition, pages 472–486. Springer.

Gurnani, A., Shah, K., Gajjar, V., Mavani, V., and Khand-

hediya, Y. (2019). Saf-bage: Salient approach for fa-

cial soft-biometric classiﬁcation-age, gender, and fa-

cial expression. In 2019 IEEE Winter Conference on

Applications of Computer Vision (WACV), pages 839–

847. IEEE.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In 2016 IEEE Con-

ference on Computer Vision and Pattern Recognition

(CVPR), pages 770–778.

Hsu, C.-Y., Lin, L.-E., and Lin, C. (2021). Age and gender

recognition with random occluded data augmentation

on facial images. Multimedia Tools and Applications.

Hu, J., Shen, L., and Sun, G. (2018). Squeeze-and-

excitation networks. In 2018 IEEE/CVF Conference

on Computer Vision and Pattern Recognition, pages

7132–7141.

Kazemi, V. and Sullivan, J. (2014). One millisecond face

alignment with an ensemble of regression trees. In

2014 IEEE Conference on Computer Vision and Pat-

tern Recognition, pages 1867–1874.

Levi, G. and Hassner, T. (2015). Age and gender classiﬁca-

tion using convolutional neural networks. In Proceed-

ings of the IEEE Conference on Computer Vision and

Pattern Recognition Workshops, pages 34–42.

Li, X., Lai, S., and Qian, X. (2021). Dbcface: Towards pure

convolutional neural network face detection. IEEE

Transactions on Circuits and Systems for Video Tech-

nology.

Li, Y., Zeng, J., Shan, S., and Chen, X. (2019). Occlusion

aware facial expression recognition using cnn with at-

tention mechanism. IEEE Transactions on Image Pro-

cessing, 28(5):2439–2450.

Liu, X. and Milanova, M. (2018). Visual attention in deep

learning: a review. In ICRA 2018.

Liu, Y., Gao, Y., and Yin, W. (2020). An improved analysis

of stochastic gradient descent with momentum. arXiv:

Optimization and Control.

Min, R., Hadid, A., and Dugelay, J. (2011). Improving the

recognition of faces occluded by facial accessories.

In 2011 IEEE International Conference on Automatic

Face Gesture Recognition (FG), pages 442–447.

Ng, C.-B., Tay, Y.-H., and Goi, B.-M. (2015). A review of

facial gender recognition. Pattern Analysis and Appli-

cations, 18(4):739–755.

Ranjan, R., Patel, V. M., and Chellappa, R. (2017). Hy-

perface: A deep multi-task learning framework for

face detection, landmark localization, pose estimation,

and gender recognition. IEEE Transactions on Pattern

Analysis and Machine Intelligence, 41(1):121–135.

Roy, A. G., Navab, N., and Wachinger, C. (2018). Concur-

rent spatial and channel squeeze & excitation in fully

convolutional networks. ArXiv, abs/1803.02579.

Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and

Chen, L.-C. (2018). Mobilenetv2: Inverted residu-

als and linear bottlenecks. In Proceedings of the IEEE

Conference on Computer Vision and Pattern Recogni-

tion, pages 4510–4520.

Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R.,

Parikh, D., and Batra, D. (2017). Grad-cam: Visual

explanations from deep networks via gradient-based

localization. In 2017 IEEE International Conference

on Computer Vision (ICCV), pages 618–626.

Simanjuntak, F. and Azzopardi, G. (2019). Fusion of cnn-

and cosﬁre-based features with application to gender

recognition from face images. In Science and Infor-

mation Conference, pages 444–458. Springer.

Wang, Q., Wu, B., Zhu, P., Li, P., Zuo, W., and Hu, Q.

(2020). Eca-net: Efﬁcient channel attention for deep

convolutional neural networks. 2020 IEEE/CVF Con-

ference on Computer Vision and Pattern Recognition

(CVPR), pages 11531–11539.

Woo, S., Park, J., Lee, J.-Y., and Kweon, I.-S. (2018).

Cbam: Convolutional block attention module. In

ECCV.

Xu, X., Saraﬁanos, N., and Kakadiaris, I. (2020). On im-

proving the generalization of face recognition in the

presence of occlusions. 2020 IEEE/CVF Conference

on Computer Vision and Pattern Recognition Work-

shops (CVPRW), pages 3470–3480.

Yang, S., Luo, P., Loy, C.-C., and Tang, X. (2016). Wider

face: A face detection benchmark. In Proceedings of

the IEEE Conference on Computer Vision and Pattern

Recognition, pages 5525–5533.

Yuan, Z. (2020). Face detection and recognition based on

visual attention mechanism guidance model in unre-

stricted posture. Scientiﬁc Programming, 2020:1–10.

yuan Wang, Z., Wang, G., Huang, B., Xiong, Z., Hong, Q.,

Wu, H., Yi, P., Jiang, K., Wang, N., Pei, Y., Chen, H.,

Miao, Y., Huang, Z., and Liang, J. (2020). Masked

face recognition dataset and application. ArXiv,

abs/2003.09093.

Zhang, W., Smith, M. L., Smith, L. N., and Farooq, A.

(2016). Gender and gaze gesture recognition for

human-computer interaction. Computer Vision and

Image Understanding, 149:32–50.

Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., and Tor-

ralba, A. (2016). Learning deep features for discrim-

inative localization. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 2921–2929.

VISAPP 2022 - 17th International Conference on Computer Vision Theory and Applications

678