Automatic Skin Tone Extraction for Visagism Applications

Diana Borza

, Adrian Darabant

and Radu Danescu

Computer Science Department, Technical University of Cluj-Napoca,

28 Memorandumului Street , 400114, Cluj Napoca, Romania

Computer Science Department, Babes Bolyai University,

58-60 Teodor Mihali Street, C333,

Cluj Napoca 400591, Romania

Keywords: Skin Tone, Color Classification, Support Vector Machine, Convolutional Neural Networks.

Abstract: In this paper we propose a skin tone classification system on three skin colors: dark, medium and light. We

work on two methods which don’t require any camera or color calibration. The first computes color

histograms in various color spaces on representative facial sliding patches that are further combined in a large

feature vector. The dimensionality of this vector is reduced using Principal Component Analysis a Support

Vector Machine determines the skin color of each region. The skin tone is extrapolated using a voting schema.

The second method uses Convolutional Neural Networks to automatically extract chromatic features from

augmented sets of facial images. Both algorithms were trained and tested on publicly available datasets. The

SVM method achieves an accuracy of 86.67%, while the CNN approach obtains an accuracy of 91.29%. The

proposed system is developed as an automatic analysis module in an optical visagism system where the skin

tone is used in an eyewear virtual try-on software that allows users to virtually try glasses on their face using

a mobile device with a camera. The system proposes only esthetically and functionally fit frames to the user,

based on some facial features –skin tone included.

1 INTRODUCTION

In modern society, physical look is an essential aspect

and people often resort to several fashion tips to

enhance their appearance. Recently, a new concept

based on the search of beauty has emerged – visagism

(Juillard, 2016). Its main goal is to ensure the perfect

harmony between one`s personality and appearance,

by using some tricks (shape and color of the

eyeglasses, hairstyle, makeup etc.) to attenuate or, on

the contrary, to highlight some features of the face.

Spectacles sales represent more than 50% of the

overall market share. In 2015 the global eyewear

market was valued 102.66 billion USD and is

continuously expanding (Grand View Research,

2016). An important step in the proposal and selling

of eyeglasses is the choice of the frame; this decision

must take into account several factors (such as the

shape of the face, skin tone, and the eye color etc.),

and opticians do not have knowledge in handling

these aspects. An automatic framework that

accurately classifies these features can assist

customers in making the appropriate choice at a much

smaller cost than training multiple employees in the

field of visagism or using fully qualified estheticians.

Soft biometrics complement the identity

information provided by traditional biometric

systems using physical and behavioral traits of the

individuals (iris and skin color, gender, gait etc.); they

are non-obtrusive, don’t require human cooperation

and can still provide valuable information. Skin tone

represents a valuable soft biometric trait.

In this paper, we propose an automatic skin tone

classification system mainly intended for the specific

use case of eyeglasses selection. Recently, several

virtual eyeglasses try-on applications have been

developed and their databases typically contain

thousands of digitized 3D frames. Physically trying a

large number of frames in reality is time challenging

and the prospective eyewear buyer often loses interest

early in the process. A virtual try-on system assisted

by an intelligent module that is able to select the

frames esthetically and physically adapted to the user

transforms this choice in a playful recreation. Each

pair of virtual eyeglasses is given a score for each

facial feature trait (gender, skin tone, hair color, shape

of the face) and its total score is computed by a

466

Borza, D., Darabant, A. and Danescu, R.

Automatic Skin Tone Extraction for Visagism Applications.

DOI: 10.5220/0006711104660473

In Proceedings of the 13th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2018) - Volume 4: VISAPP, pages

466-473

ISBN: 978-989-758-290-5

weighted average of these features and the glasses are

displayed to the used in decreasing order of their

score.

The proposed system automatically detects the

skin tone and it is integrated in this frame selection

method. For this particular use-case, the

differentiation of skin tones into three classes (dark,

medium and light) is sufficient.

Color classification is highly sensitive to

capturing devices and illumination conditions; in the

case of skin tone classification, the problem is even

harder, due to the fact the skin tones are very close

and similar, and these perturbing factors have an even

higher impact. Moreover, skin color labelling is often

found subjective even by trained practitioners

(Fitzpatrick, 1998).

We propose and compare two methods for

classifying the skin tone. The first method uses the

conventional stages of machine learning: region of

interest selection, feature extraction and

classification. Its main contributions consist in

combining and organizing multiple color-spaces into

histograms of skin patches and reducing the resulting

feature space so that only the features with high

discriminative power are kept for voting. A Support

Vector Machine (SVM) classifier is trained and used

to assign the skin color label to each skin patch.

The second method uses deep learning: the

classical stages of machine learning replaced by the

convolutional neural network (CNN) which also

learns the features which are relevant in the

classification problem.

For the training and testing steps, we have

gathered and annotated facial images from the

Internet and from different publicly available

databases.

The remainder of this paper is organized as

follows: in Section 2 we describe the state of the art

methods used for skin detection and skin color

classification and in Section 3 we detail the proposed

solution. The experimental results are discussed in

Section 4. Section 5 provides the conclusions and

directions for future work.

2 STATE OF THE ART

Most of the research conducted on skin color analysis

is focused on skin detection (Kakumanu et al., 2007)

because of its usefulness in many computer vision

tasks such as face detection and tracking (Pujol et al.,

2017).

The first attempt to create a taxonomy for skin

color was made in 1897 by Felix von Luschan who

defined a chromatic scale with 36 categories (von

Luschan, 1879). The classification was performed by

comparing the subject`s skin with painted glass tiles.

This color scale is rather problematic as it is often

inconsistent: trained practitioners give different

results to the same skin tone. Although it was largely

used in early anthropometric studies, nowadays the

von Luschan chromatic scale is replaced by novel

spectro-photometric methods (Thibodeau et al.,

1997). The Fitzpatrick scale (Fitzpatrick, 1988) is a

recognized dermatological tool for skin type

classification. This classification schema was

developed in 1975 and it uses six skin color classes to

describe sun-tanning behavior. However, this scale

needs training and is subjective.

Color is a prominent feature for image

representation and has the important advantage of

being invariant to geometrical transformations.

However, color classification proves to be a difficult

task mainly due to the influence of the illumination

conditions: a simple change in the light source, its

nature or illumination level can strongly affect the

color appearance of the object. Moreover, the

classification performance is also limited by the

quality of the image capturing devices.

Recently, with the new developments in computer

vision several works attempted to classify skin color

from images. In (Jmal et al., 2014) the skin tone is

roughly classified into two classes: light and dark. The

face region is first extracted with a general face

detector (Viola and Jones, 2001) and the skin pixels are

determined by applying some thresholds on the R, G,

B channels. To classify the skin tone, several distances

between the test frame and two reference frames are

analyzed. This method achieves 87% accuracy on a

subset of the Color FERET image database.

In (Boaventura et al., 2006) the skin color is

differentiated into three classes: dark, brown and

light, using 27 inference rules and fuzzy sets

generated from the R, G, B values of each pixel. The

method was trained and evaluated on images from the

AR dataset and images from the Internet and it

achieves a hit rate above 70%. Finer classifications

(16 skin tones) are proposed in (Harville et al., 2005)

and (Yoon et al., 2006), but these methods involve the

use of a color calibration target that contains several

predefined colors arranged in a distinctive pattern.

The calibration pattern is used for color normalization

and skin tone classification.

In the context of racial or ethnical classification

from facial images (Fu et al., 2014), some methods

use the skin tone as a cue for the race (Xie et al.,

2012), especially in the case of degraded facial

images, where other features cannot be exploited.

Automatic Skin Tone Extraction for Visagism Applications

467

Skin color tones are close to each other and

illumination changes make color even harder to

distinguish. Although, in the field of dermatology

(Fitzpatrick, 1988) and anthropometry (von Luschan,

1897) the skin color is classified at a higher

granularity level (using 6 and 36 skin tones,

respectively), for the particular applications of

visagism and soft biometrics and a simple taxonomy

with three classes (light, medium and dark) is

sufficient. More complex classification schemes are

highly subjective and pose problems even for trained

human practitioners. A model with six colors would

probably be ideal as the six tones would closely match

what we can distinguish visually amongst different

regions and human races as predominant skin colors.

However, practical studies (Boaventura et al., 2006)

show that natural, non-influenced classification of

skin colors as performed by humans would contain

only three classes: white/light, brown and black.

However, even with three classes, the classification is

subjective.

Figure 1: Example of skin patches belonging to the three

skin tone classes.

Given these aspects, in the remaining of this paper

we only consider the three above classes, based on the

idea that when a finer classification is needed it can

be derived from these three. Figure 1 shows some

skin tone examples belonging to each skin tone class.

3 PROPOSED SOLUTION

This paper presents a fully automatic skin tone

classification framework that does not require any

prior camera calibration or additional calibration

patterns. We propose and compare two methods for

classifying the skin tone into three classes: light,

medium and dark.

The first method uses support vector machine and

histograms of local image patches, while the later one

is based on convolutional neural networks.

3.1 Skin Tone Classification using

SVM

A moving window is used to compute color

histograms of local image patches in several color

spaces on a skin region from the face. The histograms

are concatenated into a single feature vector and

Principal Component Analysis (PCA) is used to

reduce its dimensionality of the feature vector.

The reduced histogram from each facial skin

patch is fed to a Support Vector Machine (SVM) to

determine the skin tone of that region, and, finally the

skin color is determined using voting. The outline of

the proposed solution is depicted in Figure 2.

The choice of the color space is critical in color

classification. Each color-space represents color

features in different ways, so that colors are more

intuitively distinguished or certain computations are

more suitable. However, none of the color-spaces can

be considered as a universal solution. In this work, we

classify the skin tone by combining the histograms of

the most commonly used color-spaces: RGB, HSV,

Lab and YCrCb.

As illustrated on Figure 3, using the three simple

skin tone classes we proposed, the problem is not

trivial as the samples cannot be clearly separated in

different color spaces. Figure 3 (a), (b) and (c)

illustrates the color distribution of the three selected

skin tones in RGB, HSV and YCrCb color spaces.

Figure 3 (d) plots the grayscale values of the three

classes in order to determine if the intensity feature

could bring any additional information to the

classification problem. Finally, in Figure 3 (e) the 3D

color points from the RGB color-space are projected

onto the 2D space using PCA by preserving only the

two axes with the most data variation.

The first step of the algorithm is face detection:

we use the popular Viola-Jones (Viola and Jones,

2001) algorithm for face localization. However, the

face region contains several non-skin pixels, such as

eyes, lips, hair etc. that could influence the

classification performance. In order to avoid these

non-relevant features, we crop the face to a region of

interest (ROI) right beneath the eyes and above the

center of the face and only the pixels in this area will

be analyzed to classify the skin tone.

The crop proportions were heuristically determi-

ned as [0.2



w, 0.3



h, 0.6



w, 0.5



h], where w

and h are the width and the height of the face region.

Image

Face

localization

& cropping

Feature

extraction

PCA

reduction

SVM

Classifier

Skin

region

histogram

DARK

MEDIUM

LIGHT

Figure 2: Flowchart of the SVM based-classification method.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

468

Figure 3: Color distribution for each skin tone. Red points represent skin pixels of dark skin tone, blue points represent skin

pixels of medium skin tone and orange points represent pixels of light skin tone.

Figure 4: Feature extraction.

Next, a moving window of size s x s (s = 21



pixels in our experiments) iterates over the ROI and

the histograms for each color component of the 4

selected color spaces are computed and concatenated.

The feature vector (if all the 4 color spaces are used)

is composed of 13 histograms (3 color-spaces and

grayscale) and has a high dimensionality (i.e. 12 

256 = 3328 bins). In such cases the training problem

can be difficult and the high dimensional input space

can “confuse” the learning algorithm and lead to

over-fitting. Therefore, we apply a dimensionality

reduction pre-processing step (Principal Component

Analysis) in order to increase the robustness towards

illumination conditions and to reduce the time

complexity of the classifier. Using PCA, the original

high dimensional input space is reduced (with some

data loss, but retaining as much variance as possible)

to a lower dimensional space determined by the

highest eigenvectors. We retain the first principal

components such that p = 98% of the variance of the

data is preserved; the number of retained principal

components is computed on the training set and once

computed it is fixed. This value was determined

heuristically through trial and error experiments and

is fixed once determined.

Figure 4 illustrates the feature extraction process.

The reduced feature vector resulted from each

position of the sliding window is fed to a Support

Vector Machine classifier in order to obtain the skin

tone of the region. A pre-processing step is applied,

by scaling the input such that each feature from the

training set has zero mean and unit variance.

SVM classifiers are supervised learning

algorithms originally developed for binary linear

classification problems. To adapt the classifier to our

3 class classification problem, we used “one versus

one” approach: n



(n - 1)/2 classifiers are constructed

(each SVM must learn to distinguish between two

classes) and at prediction time a voting scheme is

applied.

We use the Radial basis function (RBF) kernel or

Gaussian kernel:











 













 (1)

Finally, the skin tone classification is the

dominant skin tone within the selected ROI computed

Automatic Skin Tone Extraction for Visagism Applications

469

by majority voting. For each sliding window position,

the classifier gives a skin label and the final skin tone

is selected as the class that occurs most often.

3.2 Skin Tone Classification using CNN

Recently, in the field of computer vision CNN have

achieved impressive results in a variety of

classification tasks. The CNN structure is based on

mechanisms found in the visual cortex of the

biological brain. The neurons in a CNN are arranged

into 3 dimensions: (width, height and depth) and the

neurons within a layer are only connected to a small

region of the previous layer, the receptive field of the

neuron. Typically, CNNs have three types of basic

layers: the convolutional layer (CONV) – followed by

a non-linearity, e.g. Rectified Linear Unit (ReLU) –,

the pooling layer and the fully connected (FC) layer.

The VGG (Simonyan and Zisserman, 2014)

network modified the traditional CNN architecture by

adding more convolutional layers (19 layers for the

VGG-19 network) and reducing the size of the

convolutional layers (to 3x3 convolutional filters with

1 stride in all layers). The structure of the VGG-19

network is the following: 2



conv3_64, maxpool,



conv3_128, maxpool, 4



conv3_256, maxpool,



conv3_512, maxpool, 4



conv3_512, maxpool, 2



FC_4096, FC_1000 and softmax; where convs_d is a

convolutional layer of size s and depth d, maxpool is

a max-pooling layer and FC_n is a fully connected

layer with n neurons.

We finely tuned a VGG-19 network which was

previously trained for gender recognition from facial

images (Rothe et al., 2016). The network was trained

from scratch using more than 200.000 images and

uses RGB, 224



224 images as input.

The only constraint we apply on a new image is

that it contains a face which can be detected by a

general face detector (Viola and Jones, 2001). The

face region is further enlarged with 40% horizontally

and vertically and this region is fed to the

convolutional neural network. The training is

performed using the rmsprop optimizer; the batch

size was set to 128 and the rmsprop momentum to

0.9; the learning rate is initially set to 10

-2

and then

exponentially decreased after each to epochs.

To tune the network for the skin tone

classification problem, we removed the last two fully-

connected layers of the trained CNN and the

remaining part of the convolutional neural network is

treated as a fixed feature extractor. Finally, a linear

classifier (softmax) is trained for the skin tone

classification problem using the features previously

learned by the CNN.

4 EXPERIMENTAL RESULTS

AND DISCUSSIONS

4.1 Training and Tests Datasets

To create the training dataset, we have fused together

images from multiple face databases ((Caltech,

1999), (Ma et al., 2015), (Minear and Park, 2004),

(Thomaz and Giraldi, 2010)) and labelled the images

according to the skin color of the subject. The Caltech

(Caltech, 1999) face database contains 450 outdoor

and indoor images captured in uncontrolled lighting

conditions. The Chicago Face Database (Ma et al.,

2015) comprises more than 2000 images acquired in

controlled environments. The Minear-Park database

(Minear and Park, 2004) contains 576 facial images

captured in natural lightning conditions and the

Brazilian face database (Thomaz and Giraldi, 2010)

is composed of 2800 frontal facial images captured

against a homogenous background.

In addition, the training dataset was enlarged with

images of celebrities captured in unconstrained

conditions. For this, we extracted the names of

celebrities with different skin tones (www.listal.com)

and we crawled Internet face images of the selected

celebrities.

To determine the actual skin tone, each image

sample was independently annotated by three

different persons and the ground truth was determined

by merging the results from the independent

annotations (majority voting). First, the annotators

had a training with a visagist expert who instructed

them with the rules two follow in the annotation

process. During the training, they also annotated

together with the visagist a subset of the images and

discussed how to handle boundary cases and

uncertainties. After merging the human labeling

results, we observed that most of the annotation

inconsistencies appeared between the medium and

the dark skin tones.

We have used several augmentation techniques:

contrast stretching and brightness enhancement. As

some of the datasets are captured in controlled

scenarios ((Ma et al., 2015), (Minear and Park,

2004)), these augmentation techniques will make

learning algorithm more robust to illumination

conditions.

We have extracted subsets of each of the four

databases such that the distribution of the skin tone

classes is approximately even. The final training

dataset consists of 8952 images. The classifier was

evaluated on 999 images that were not used in the

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

470

training process; the test dataset is balanced: we use

333 images for each class.

4.2 SVM Classification Results

We trained the SVM classifier using different color

features: first we have used the histograms of the four

selected color spaces (RGB, HSV, Lab and YCrCb)

independently, next we have also added the grayscale

histogram (when not redundant) to each color space,

and finally we have combined multiple color-spaces.

Table 1 shows the classification accuracies for the

different color features used.

Table 1: Classification performance using different color

spaces.

Color channels

Accuracy

R, G, B

74.65%

H, S, V

83.30%

L, a, b

72.46%

Y, Cr, Cb

83.30%

L, a, b, H, S, V

85.18%

L, a, b, Grayscale

84.19%

L, a, b, H, S, R, G, B, Y, Cr, Cb

86.67%

Table 2: Confusion matrix for the VGG-19 convolutional

neural network.

Predicted Value

Dark

Light

Medium

Actual

value

Dark

271

Light

302

Medium

299

The best results are obtained by combining

multiple color-spaces (L, a, b, H, S, R, G, B, Y, Cr,

Cb): 86.67% accuracy. The confusion matrix for this

experimental setup is reported in Table 2.

4.3 CNN classification results

The classification report using the convolutional

neural network is reported in Table 3.

Table 3: Classification report for the VGG-19

convolutional neural network.

Class

Precision

Recall

F1-Score

Dark

0.9410

0.9099

0.9252

Light

0.9868

0.8979

0.9403

Medium

0.8289

0.9309

0.8769

The overall accuracy obtained using the VGG-19

network is 91.29%. By using a convolutional neural

network to classify the skin tone we obtained a

performance boost by 4.62%.

Table 4 shows the confusion matrix for the CNN

classification. From the confusion matrix, it can be

noticed that the majority of “confusions” occurred

between medium-light skin tones and dark-medium

skin tones. This behavior is very similar to what we

observed in the annotations of the ground truth by the

three human labelers. As opposed to the SVM skin

classification approach, it can be noticed that there no

confusions between the Dark-Light and Light-Dark

classes.

Table 4: Confusion matrix for the VGG-19 convolutional

neural network.

Predicted Value

Dark

Light

Medium

Actual

value

Dark

303

Light

299

Medium

310

4.4 Comparison to State of the Art

Next, in this section we compare the proposed

solution with other state of the art works that tackle

the problem of skin tone classification.

We obtain the best using the VGG-19

convolutional neural network (accuracy 91.29%).

Not all the proposed methods classify the skin tones

at the same granularity level. In (Jmal et al., 2014) the

obtained classification accuracy is 87% but the skin

tone is distinguished into only dark and light. In

(Boaventura et al., 2006) the skin color is divided into

the same 3 classes used in the proposed solution:

light, medium and black, but the obtained accuracy is

only 70%. However, the method (Boaventura et al.,

2006) was tested on images captured in the authors’

laboratory and on a subset of the AR dataset, so the

testing benchmark is not publically available. Finally,

in (Harville et al., 2005) the authors make use of a

color calibration pattern that must be held by the

subjects in each test image. The color calibration

pattern arranges the primary and secondary colors,

and 16 patches representative of the range of human

skin color into a known pattern. The main

disadvantage of that method is that is assumes a

controlled image capturing scenario (the user must

hold the color chart).

Our method does not impose any restrictions of

the image capturing scenario and attains an accuracy

rate of 91.29%.

Some examples of correctly classified samples are

depicted in Figure 5.

Automatic Skin Tone Extraction for Visagism Applications

471

Figure 5: Examples of correctly classified images: (a) dark

as dark, (b) medium as medium, (c) dark as dark.

Figure 6 shows some examples of incorrectly

classified images.

Figure 6: Examples of incorrectly classified images:

(a) dark as medium, (b) medium as dark, (c) light as

medium.

4.5 Applications

The proposed approach is intended for a facial

attribute analysis system used in virtual eyeglasses

try-on (Figure 7).

Figure 7: Outline of the eyewear-proposal system.

First, a facial image of the subject is captured and

the system (Facial attributes extraction module)

automatically determines the skin tone (and other

demographical attributes: gender, age, eye color etc.).

Based on these attributes, the Frame Selection

module queries the 3D eyeglasses database and

selects the accessories that are esthetically and

functionally in harmony with the user`s face. Each 3D

eyeglasses pair was previously annotated by a

specialized visagist/esthetician with a score for each

facial attribute and only the eyeglasses with the

highest scores are displayed to the user. Typically, the

3D eyeglasses dataset contains more than several

thousand eyeglasses models.

More specifically, the system is implemented in

Objective-C for an iPad application: first, a picture of

the user is analyzed in order to determine the skin tone

(and other demographical attributes) and the most

appropriate pairs of glasses are selected from the

database. Next, the system starts to track the user`s

face and uses augmented reality to place the selected

eyeglasses on the subject`s face. The system was also

tested on images captured in this scenario.

Of course, other applications can be envisioned:

the data extracted by the Facial attributes extraction

module can be used, for instance, to suggest the

appropriate make-up or hair color, to estimate skin

tone distribution over populations over geographical

areas etc.

5 CONCLUSIONS

This paper presented an automatic skin tone

classification system that doesn’t require any

additional color patterns or prior camera color

calibration. We proposed and compared two methods

for classifying the skin tone in facial input images.

The first method uses conventional machine

learning techniques: histogram of various skin

patches from the face and a SVM to determine the

skin tone. First, the face is localized in the input image

and it is cropped to a region that is most likely to

contain only skin pixels. Next, a window slides over

this region and color histograms in different color

spaces are computed and concatenated for each

window position. The feature vector is reduced using

PCA and a SVM classifier determines the skin tone

of the window. The skin tone is determined using a

simple voting procedure on the result of each

histogram patch from the region of interest.

The later method uses convolutional neural

networks, which also learn the relevant chromatic

features for the skin tone classification problem. We

finely tuned a neural network which was previously

trained on the problem of gender detection from facial

images. The system was trained and tested on images

from four publicly available datasets and from images

crawled from the Internet.

As a future work, a more complex method to

determine the skin pixels within the face area is

envisioned. In addition, we plan to integrate the

current method with a full visagism analysis system

that also determines the eyes color, the hair color and

the face shape.

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

472

ACKNOWLEDGEMENTS

This work was supported by the MULTIFACE grant

(Multifocal System for Real Time Tracking of

Dynamic Facial and Body Features) of the Romanian

National Authority for Scientific Research, CNDI–

UEFISCDI, Project code: PN-II-RU-TE-2014-4-

1746.

REFERENCES

Boaventura I. A. G., Volpe V. M., da Silva I. N., Gonzaga

A., 2006. Fuzzy classiﬁcation of human skin color in

color images. In IEEE International Conference on

Systems, Man and Cybernetics, 5071– 5075.

Caltech University, 1999. Caltech Face Database -

http://www.vision. caltech. edu. Image Datasets/faces.

Fitzpatrick, T. B., 1988. The validity and practicality of

sun-reactive skin types I through VI. Archives of

dermatology, 124(6), pp.869-871.

Fu, S., He, H. and Hou, Z. G., 2014. Learning race from

face: A survey. IEEE transactions on pattern analysis

and machine intelligence, 36(12), pp. 2483-2509.

Grand View Research, 2016. Eyewear Market Analysis by

Product (Contact Lenses, Spectacles, Plano Sunglasses)

And Segment Forecasts To 2024, Available at:

http://www.grandviewresearch.com/industry-analysis/

eyewear-industry, [Accessed: 13.04.2017]

Harville, M., Baker, H., Bhatti, N. and Susstrunk, S., 2005,

September. Consistent image-based measurement and

classification of skin color. In Image Processing, 2005.

ICIP 2005. IEEE International Conference on (Vol. 2,

pp. II-374). IEEE.

Jmal M., Souidene W., Attia R., Youssef A., 2014.

Classiﬁcation of human skin color and its application to

face recognition. In: The Sixth International Conferen-

ces on Advances in Multimedia MMEDIA 2014.

Juillard C., 2016. Brochure Méthode C. Juillard 2016,

Available at: http://www.visagisme.net/Brochure-

Methode-C-JUILLARD-2016.html, [Accessed: 30.07.

2017]

Kakumanu, P., Makrogiannis, S. and Bourbakis, N., 2007.

A survey of skin-color modeling and detection

methods. Pattern recognition, 40(3), pp.1106-1122.

von Luschan B., 1897. Beitrage zur Volkerkunde der

deutschen Schutzgebiete. D. Reimer.

Ma, D. S., Correll, J. and Wittenbrink, B., 2015. The

Chicago face database: A free stimulus set of faces and

norming data. Behavior research methods, 47(4),

pp.1122-1135.

Minear, M. and Park, D. C., 2004. A lifespan database of

adult facial stimuli. Behavior Research Methods,

Instruments, & Computers, 36(4), pp.630-633.

Pujol, F. A., Pujol, M., Jimeno-Morenilla, A. and Pujol, M.

J., 2017. Face detection based on skin color segmenta-

tion using fuzzy entropy. Entropy, 19(1), p.26.

Rothe R., Timofte R., van Gool L.. 2016, "Deep expectation

of real and apparent age from a single image without

facial landmarks." International Journal of Computer

Vision (2016): 1-14.

Simonyan, K., Zisserman, A. 2014. Very deep

convolutional networks for large-scale image

recognition. arXiv preprint arXiv:1409.1556.

Thibodeau, E. A. and D'Ambrosio, J. A., 1997.

Measurement of lip and skin pigmentation using

reflectance spectrophotometry. European journal of

oral sciences, 105(4), pp.373-375.

Thomaz, C. E. and Giraldi, G. A., 2010. A new ranking

method for principal components analysis and its

application to face image analysis. Image and Vision

Computing, 28(6), pp.902-913.

Viola, P. and Jones, M., 2001. Rapid object detection using

a boosted cascade of simple features. In Computer

Vision and Pattern Recognition, 2001. CVPR 2001.

Proceedings of the 2001 IEEE Computer Society

Conference on (Vol. 1, pp. I-I). IEEE.

Yoon, S., Harville, M., Baker, H. and Bhatii, N., 2006,

October. Automatic skin pixel selection and skin color

classification. In Image Processing, 2006 IEEE

International Conference on (pp. 941-944). IEEE.

Xie, Y., Luu, K. and Savvides, M., 2012. A robust approach

to facial ethnicity classification on large scale face

databases. In Biometrics: Theory, Applications and

Systems (BTAS), 2012 IEEE Fifth International

Conference on (pp. 143-149).

Automatic Skin Tone Extraction for Visagism Applications

473