Combination of Texture and Geometric Features for Age Estimation in

Face Images

Marcos Vinicius Mussel Cirne and Helio Pedrini

Institute of Computing, University of Campinas, Campinas, SP, 13083-852, Brazil

Keywords:

Age Estimation, Image Analysis, Texture, Geometric Descriptor.

Abstract:

Automatic age estimation from facial images has recently received an increasing interest due to a variety

of applications, such as surveillance, human-computer interaction, forensics, and recommendation systems.

Despite such advances, age estimation remains an open problem due to several challenges associated with

the aging process. In this work, we develop and analyze an automatic age estimation method from face

images based on a combination of textural and geometric features. Experiments are conducted on the Adience

dataset (Adience Benchmark, 2017; Eidinger et al., 2014), a large known benchmark used to evaluate both age

and gender classiﬁcation approaches.

1 INTRODUCTION

Biometric systems commonly employ a number

of distinctive measurable human characteristics to

recognize individuals, such as face, ﬁngerprint,

palm print, deoxyribonucleic acid (DNA) and

iris (Paulo Carlos et al., 2015; Pinto et al., 2015;

Silva Pinto et al., 2012; Menotti et al., 2015; Menotti

et al., 2015; Silva et al., 2015; Assis Angeloni and

Pedrini, 2016). Soft biometrics (Jain et al., 2004)

refer to metrics related to physical or behavioral

human characteristics, such as gender, age, hair

color, height and weight, which are complementary

to primary biometric identiﬁers. The combination

of primary and soft biometric characteristics can

signiﬁcantly improve the performance of person

recognition in surveillance systems.

The problem of age estimation from face im-

ages (Choi et al., 2011; Huerta et al., 2014; Lani-

tis et al., 2002; Liu et al., 2015; Ren and Li, 2014;

Thukral et al., 2012; Geng et al., 2013) is very chal-

lenging due to the inherently complex nature of the

aging process, high variability within a same age in-

terval, personal characteristics of each individual, as

well as difﬁculties in collecting large datasets derived

from chronological images from the same individuals.

Despite its large applicability in several knowl-

edge domains, there is still relatively little research

on age estimation compared to other facial analysis

topics, such as face and iris recognition. Examples

of applications of age estimation from facial images

include forensics, surveillance, human-computer in-

teraction, and recommendation systems.

In this work, we propose and evaluate a novel age

estimation approach from facial images based on a

combination of textural and geometric features.

Experiments are conducted on the Adience

dataset (Adience Benchmark, 2017; Eidinger et al.,

2014), which is a well known benchmark used to

evaluate age estimation approaches.

The remainder of this paper is organized as fol-

lows. A literature review about age estimation ap-

proaches is shown in Section 2. The proposed age es-

timation method is detailed in Section 3. Experimen-

tal results are presented and discussed in Section 4.

Finally, some ﬁnal remarks and directions for future

work are included in Section 5.

2 BACKGROUND

Due to a variety of applications, there has been an

increasing interest of the scientiﬁc community in in-

vestigating the automatic age estimation from facial

images.

Most of the age estimation approaches available

in the literature are based on feature extraction and

learning algorithms. A pioneer work on age estima-

tion from facial images was proposed by Kwon and

Lobo (Kwon and da Vitoria Lobo, 1994), which was

based on an anthropomorphic model to differentiate

Cirne, M. and Pedrini, H.

Combination of Texture and Geometric Features for Age Estimation in Face Images.

DOI: 10.5220/0006625503950401

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

395-401

ISBN: 978-989-758-290-5

395

three age groups: babies, young adults and senior

adults. In the same work, they also proposed a wrin-

kle detection method based on snakelets.

Lanitis et al. (Lanitis et al., 2002) employed ac-

tive appearance models (AAM) for age estimation by

deﬁning an ageing function. Chang et al. (Chang

et al., 2011) also used AAM to estimate ordinal hy-

perplanes and ranks them according to age intervals.

Geng et al. (Geng et al., 2013) explored an ageing

pattern subspace model to extract and process feature

vector for age estimation.

Fu and Huang (Fu and Huang, 2008) developed a

manifold embedding approach to the age estimation

problem, whose purpose is to ﬁnd a low-dimensional

representation in the embedded subspace and cap-

ture geometric structure and data distribution. Wu et

al. (Wu et al., 2012) explored a Grassmann manifold

to model the facial shapes and considered the age es-

timation as regression and classiﬁcation problems on

this representation.

Appearance models have also been explored by

several authors for age estimation purpose. Gao

and Ai (Gao and Ai, 2009) used a texture descrip-

tor based on Gabor ﬁlter to estimate age. Guo and

Guowang (Guo and Mu, 2014), Guo et al. (Guo et al.,

2009) and Weng et al. (Weng et al., 2013) employed

biologically inspired features for age estimation from

face images.

Hayashi et al. (Hayashi et al., 2002) developed a

method for age estimation based on wrinkle texture

and color information extracted from facial images.

Shape and size of the facial parts were used to predict

the age.

Iga et al. (Iga et al., 2003) described extraction

functions from facial candidate regions using color

information and parts of the face. Age was estimated

based on SVM classiﬁers and a voting scheme from

the extracted features.

Suo et al. (Suo et al., 2010) proposed a composi-

tional and dynamic model for age estimation, where a

hierarchical and-or graph was used to represent faces

in each age group. A Markov process was employed

to parse the graph representation.

Kawano et al. (Kawano et al., 2005) described a

four-directional feature based on multiple parts of the

face, such as nose, lip, jaw, eyes. Linear discriminant

analysis was employed to recognize these image re-

gions.

Luu et al. (Luu et al., 2011) developed an appear-

ance model based on contourlet transforms to locate

facial landmarks. Local and holistic texture informa-

tion was explored for facial age estimation.

For further details on age estimation models and

algorithms, the reader can refer to surveys by Dhimar

and Mistree (Dhimar and Mistree, 2016) and Fu et

al. (Fu et al., 2010).

3 PROPOSED METHOD

The proposed approach to the age classiﬁcation prob-

lem is divided into two major steps: pre-processing

stage and cross-validation stage. Figure 1 shows an

overview of these stages.

In the pre-processing stage, a dataset containing

several images of people of different ages (or intervals

of ages) is analyzed. After that, the feature extrac-

tion process is started. For each image of the dataset,

several image descriptors are extracted, providing a

“descriptor database” that allows the use of different

combinations of descriptors.

For this work, both image and geometric features

were taken into account for estimating ages from im-

ages. Concerning image features, two descriptors

were extracted from the images: the ﬁrst one is the

Histogram of Oriented Gradients (HOG) (Dalal and

Triggs, 2005), a very popular and robust descrip-

tor used for detection and recognition of objects and

faces (D

eniz et al., 2011; Felzenszwalb et al., 2010;

Suard et al., 2006; Zhu et al., 2006).

First, each image was downsampled to a size of

64×64 pixels in order to produce a more compact de-

scriptor. To calculate the HOG descriptor, the images

were initially split into 8× 8 cells. Then, for the block

normalization process, a block size of 16 × 16 with a

stride of 8 × 8 was used, thus producing 7 horizon-

tal and 7 vertical positions, resulting 49 positions. By

deﬁning a histogram size of 9 bins, since each block

encompasses 4 cells and each cell provides a single

histogram, each step of the block normalization pro-

cess creates a feature vector of 36 dimensions. Hence,

the ﬁnal HOG descriptor has a size of 36 × 49 = 1764

dimensions.

The second image descriptor used in our method

was the CENTRIST (CENsus TRansform hIS-

Togram) (Wu and Rehg, 2011), an evolution of

the LBP-texture descriptor (Ojala et al., 2002) that

encodes the structural properties of an image at

the same time it has a high level of robustness to

illumination variations. Here, for each pixel x

an grayscale image, its intensity is compared against

its 8 neighbors x

, where p = 0, 1, ...7, producing a

bit string B

= b

...b

. For each comparison, if

≥ x

, then we have b

= 1. Otherwise, b

= 0.

The ﬁnal bit string is then converted to a base-10

value in the range [0, 255].

Equation 1 shows the pattern used in this pro-

cess, along with an example. Once this procedure is

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

396

PRE-PROCESSING

STAGE

CROSS-VALIDATION

STAGE

FACE IMAGE

DATASET

FEATURE

GENERATION

DESCRIPTOR

CONSTRUCTION

TRAINING

SPLIT

TEST

SPLIT

CLASSIFIER

TRAINING

TRAINED

MODEL

PARAMETER

TUNING

FINAL

ACCURACY

DESCRIPTOR 1

DESCRIPTOR 2

DESCRIPTOR N

CONCATENATION

OF DESCRIPTORS

Figure 1: General scheme of our approach to age classiﬁcation using combinations of image descriptors and machine learning

techniques.

done for all pixels, a histogram of 256 bins is com-

puted from the computed values, which forms the ﬁ-

nal CENTRIST descriptor.









⇒





40 160 120

180 100 80

100 20 80





⇒ B

= 01110100

= 116

(1)

Besides image features, a geometric descriptor

was also included in the set of descriptors. Given a

pre-trained face shape model and an image from the

age dataset, a face detection algorithm is run to ﬁnd

both the region that covers the face and, if a face is

found, the facial landmarks that correspond to impor-

tant facial features (such as eyes, nose, mouth, eye-

brows) according to the shape model.

In this work, the algorithm and the model pro-

vided by the Dlib toolkit (Dlib Toolkit, 2017) were

used in this process. The algorithm for face de-

tection is an implementation based on the work by

Kazemi and Sullivan (Kazemi and Sullivan, 2014),

which uses an ensemble of regression trees to esti-

mate facial landmarks from a sparse subset of pixel

intensities in a fast and efﬁcient way.

The shape model used in this process, along with

the localization and indexing of the facial landmarks,

is shown in Figure 2. Based on this model, the face

detection algorithm extracts the exact localization of

the pixels that correspond to each of the 68 points of

the model. Once the set of points is obtained, a ge-

ometric descriptor is constructed by taking the Eu-

clidean distances between all pairs of points, which

gives a total of





= 2278 distances.

Figure 2: Facial landmarks of the face shape model used in

Dlib’s face detection algorithm (Dlib Toolkit, 2017).

After extracting several image descriptors, the de-

scriptor construction process is started by creating dif-

ferent concatenations of descriptors that will repre-

Combination of Texture and Geometric Features for Age Estimation in Face Images

397

sent each image from a dataset. Since the features of

each descriptor are from different natures and sizes,

the values of the ﬁnal descriptor are rescaled so that

they have zero mean and unit variance. In this work,

7 different combinations of descriptors were used,

which will be detailed later on Section 4.

Once the pre-processing stage is ﬁnished, the

dataset is split into training and test sets using the

k-fold cross-validation. In this procedure, the entire

dataset is split into k parts such that one of these parts

is left apart for the tests and the k − 1 remaining ones

are used for training a classiﬁer. Later, the test set is

used to evaluate the accuracy of the trained model and

the entire process is repeated for the other folds such

that each fold is used once as the test set, which gives

a total of k iterations. Finally, the mean accuracy and

the standard error of the trained model are calculated

from the accuracy rates obtained in each iteration of

the cross-validation process.

For this stage of the proposed method, the multi-

class version of the Support Vector Machines (SVM)

classiﬁer (Cortes and Vapnik, 1995; Hearst et al.,

1998) with an RBF kernel was used. Other possi-

bilities of classiﬁers (Alpaydin, 2014; Bishop, 2006;

Murphy, 2012) were also considered, such as K-

Nearest Neighbors, Logistic Regression and Random

Forests, but the SVM produced the best results among

all those options.

4 EXPERIMENTAL RESULTS

The proposed method was implemented using Scikit-

Learn (Scikit-Learn Machine Learning in Python,

2017), an open source Python library that contains

several tools for machine learning, image processing,

data mining and data analysis. From the set of de-

scriptors detailed in Section 3 (HOG, CENTRIST and

the geometric descriptor), all possible combinations

among them were evaluated.

The tests were conducted on the Adience

dataset (Adience Benchmark, 2017; Eidinger et al.,

2014), created by the Open University of Israel (OUI)

to facilitate the study of both age and gender classiﬁ-

cation problems. The collection contains over 26,580

images of 2,284 subjects from 8 different age groups,

from newborns to old-aged people. All photos were

taken with several variations in appearance, posing,

lighting, background and facial expressions. Besides

the original images, a special version containing

cropped and aligned face images is also available.

In addition, the dataset provides information for a

5-fold cross-validation procedure, listing the images

that make part of each split. Figure 3 shows some

Figure 3: Examples of images from the Adience dataset (Ei-

dinger et al., 2014).

examples from the dataset.

For this work, only frontal images from the

aligned version of the dataset were used, since the

face detection algorithm does not perform very well

with rotated faces (i.e., face images that are out of the

±5

◦

range of yaw angle), making the computation

of the geometric descriptor difﬁcult for these cases.

After generating all descriptors for the remaining

images, the dataset was reduced to a total of 11,437

images. A complete list of the amount of images in

each fold for each age interval can be seen in Table 1.

The evaluation process is the same adopted by Ei-

dinger et al. (Eidinger et al., 2014). Here, two dif-

ferent accuracies are calculated: the exact accuracy,

which computes the groups that were correctly pre-

dicted, and the 1-off accuracy, which also considers

errors of one age group as correct predictions (e.g.:

for a face image whose correct class is “15-20”, “8-

12” and “25-32” predictions are also regarded as cor-

rect). Then, the mean accuracies of all 5 folds of the

database and their respective standard errors are cal-

culated. Table 2 shows the results for all combinations

of descriptors, where the boldface values represent the

best results obtained by our approach.

From the table, it can be seen that the three de-

scriptors combined produced the best results among

all possibilities for both exact and 1-off accuracy

rates. Analyzing each descriptor separately, the HOG

descriptor provided the highest accuracy rates. More-

over, when concatenating a different descriptor to an

existing one, both accuracies are increased. In other

words, the descriptor formed by the combination of

the three descriptors used in this work performs bet-

ter than all combinations of two descriptors, which

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

398

Table 1: Amounts of aligned frontal face images from the Adience dataset (Adience Benchmark, 2017; Eidinger et al., 2014)

for each fold and age interval.

Fold #

Age Intervals

0-2 4-6 8-12 15-20 25-32 38-43 48-53 60+ Total

Fold 1 642 354 153 105 1062 381 158 88 2943

Fold 2 155 326 414 355 415 288 84 92 2129

Fold 3 608 256 335 157 481 195 70 113 2215

Fold 4 108 195 389 300 570 312 65 77 2016

Fold 5 195 381 212 142 573 305 152 174 2134

Total 1708 1512 1503 1059 3101 1481 529 544 11437

Table 2: List of mean exact and 1-off accuracies for different combinations of image descriptors for the Adience dataset (Adi-

ence Benchmark, 2017; Eidinger et al., 2014).

Descriptor(s) Exact Acc. (%) 1-Off Acc. (%)

CENTRIST 30.38 ± 6.71 54.84 ± 5.19

Geometric 41.10 ± 4.76 76.55 ± 2.09

HOG 43.69 ± 5.70 77.03 ± 2.09

CENTRIST + Geometric 43.22 ± 5.67 78.51 ± 2.90

CENTRIST + HOG 45.03 ± 7.00 78.08 ± 2.66

Geometric + HOG 45.99 ± 6.07 81.07 ± 2.26

CENTRIST + Geometric + HOG 46.70 ± 6.56 81.80 ± 2.23

LBP + FBLBP (Eidinger et al., 2014) 44.5 ± 2.3 80.7 ± 1.1

LBP + FBLBP + Dropout-SVM (Eidinger et al., 2014) 45.1 ± 2.6 79.5 ± 1.4

in turn perform better than using only one descrip-

tor, except for the exact accuracy in the HOG versus

CENTRIST + Geometric case.

Considering the maximum accuracies obtained,

these results are slightly better than the ones achieved

by Eidinger et al. (Eidinger et al., 2014), who ob-

tained accuracies of 44.5 ± 2.3 and 80.7 ± 1.1 for the

exact and 1-off cases, respectively, only using vari-

ations of the LBP descriptor and linear SVM. How-

ever, they managed to improve the exact accuracy

to 45.1 ± 2.6 by using a variation of SVM called

dropout-SVM, inspired by the concept of dropout

from deep neural networks (Krizhevsky et al., 2012).

Therefore, our approach has great space for improve-

ments, not only in the choice of image descriptors, but

also in the machine learning process as a whole.

5 CONCLUSIONS AND FUTURE

WORK

This work presented an approach to age estimation

from face images using image descriptors that encom-

pass both texture and geometric features. The Adi-

ence dataset was used as the case study and trained

with an SVM classiﬁer with an RBF kernel.

Tests were conducted with several combinations

of descriptors, also observing how the addition of a

speciﬁc descriptor affects the mean accuracies. Re-

sults showed that our approach is comparable to the

LBP-feature based approach (Eidinger et al., 2014),

even though a different classiﬁer was used to achieve

the best results.

Future directions for this work include: tests

with new sets of descriptors (including a reﬁnement

of the geometric descriptor), use of different age

datasets and benchmarks and parameter optimization

for the machine learning procedure, application of

deep learning techniques to improve the quality of the

results, along with an evaluation of other types and

variations of classiﬁers.

ACKNOWLEDGMENTS

The authors are thankful to S

ao Paulo Research

Foundation (grants FAPESP #2017/12646-3 and

Combination of Texture and Geometric Features for Age Estimation in Face Images

399

#2014/12236-1) and National Council for Scien-

tiﬁc and Technological Development (grant CNPq

#305169/2015-7) for their ﬁnancial support.

REFERENCES

Adience Benchmark (2017). Unﬁltered Faces for Gen-

der and Age Classiﬁcation. http://www.openu.ac.il/

home/hassner/Adience/data.html.

Alpaydin, E. (2014). Introduction to Machine Learning.

MIT Press.

Assis Angeloni, M. and Pedrini, H. (2016). Part-based Rep-

resentation and Classiﬁcation for Face Recognition.

In IEEE International Conference on Systems, Man,

and Cybernetics, pages 2900–2905, Budapest, Hun-

gary. IEEE.

Bishop, C. M. (2006). Pattern Recognition and Machine

Learning. Springer.

Chang, K.-Y., Chen, C.-S., and Hung, Y.-P. (2011). Ordinal

Hyperplanes Ranker with Cost Sensitivities for Age

Estimation. In IEEE Conference on Computer Vision

and Pattern Recognition, pages 585–592. IEEE.

Choi, S. E., Lee, Y. J., Lee, S. J., Park, K. R., and Kim, J.

(2011). Age Estimation using a Hierarchical Classiﬁer

based on Global and Local Facial Features. Pattern

Recognition, 44(6):1262–1281.

Cortes, C. and Vapnik, V. (1995). Support-Vector Networks.

Machine Learning, 20(3):273–297.

Dalal, N. and Triggs, B. (2005). Histograms of Oriented

Gradients for Human Detection. In IEEE Computer

Society Conference on Computer Vision and Pattern

Recognition, volume 1, pages 886–893.

eniz, O., Bueno, G., Salido, J., and la Torre, F. D. (2011).

Face recognition using Histograms of Oriented Gradi-

ents. Pattern Recognition Letters, 32(12):1598–1603.

Dhimar, T. and Mistree, K. (2016). Feature Extraction for

Facial Age Estimation: A Survey. In International

Conference on Wireless Communications, Signal Pro-

cessing and Networking, pages 2243–2248.

Dlib Toolkit (2017). http://dlib.net.

Eidinger, E., Enbar, R., and Hassner, T. (2014). Age

and Gender Estimation of Unﬁltered Faces. IEEE

Transactions on Information Forensics and Security,

9(12):2170–2179.

Felzenszwalb, P. F., Girshick, R. B., McAllester, D., and

Ramanan, D. (2010). Object Detection with Discrim-

inatively Trained Part-Based Models. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

32(9):1627–1645.

Fu, Y., Guo, G., and Huang, T. S. (2010). Age Synthesis

and Estimation via Faces: A Survey. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

32(11):1955–1976.

Fu, Y. and Huang, T. S. (2008). Human Age Estimation

with Regression on Discriminative Aging Manifold.

IEEE Transactions on Multimedia, 10(4):578–584.

Gao, F. and Ai, H. (2009). Face Age Classiﬁcation on Con-

sumer Images with Gabor Feature and Fuzzy LDA

Method. In International Conference on Biometrics,

pages 132–141. Springer.

Geng, X., Yin, C., and Zhou, Z.-H. (2013). Facial Age Es-

timation by Learning from Label Distributions. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, 35(10):2401–2412.

Guo, G. and Mu, G. (2014). A Framework for Joint Es-

timation of Age, Gender and Ethnicity on a Large

Database. Image and Vision Computing, 32(10):761–

770.

Guo, G., Mu, G., Fu, Y., and Huang, T. S. (2009). Human

Age Estimation using Bio-inspired Features. In IEEE

Conference on Computer Vision and Pattern Recogni-

tion, pages 112–119. IEEE.

Hayashi, J.-i., Yasumoto, M., Ito, H., and Koshimizu, H.

(2002). Age and Gender Estimation based on Wrinkle

Texture and Color of Facial Images. In 16th Interna-

tional Conference on Pattern Recognition, volume 1,

pages 405–408. IEEE.

Hearst, M. A., Dumais, S. T., Osuna, E., Platt, J.,

and Scholkopf, B. (1998). Support Vector Ma-

chines. IEEE Intelligent Systems and their Applica-

tions, 13(4):18–28.

Huerta, I., Fern

andez, C., and Prati, A. (2014). Facial Age

Estimation through the Fusion of Texture and Local

Appearance Descriptors. In European Conference on

Computer Vision, pages 667–681. Springer Interna-

tional Publishing.

Iga, R., Izumi, K., Hayashi, H., Fukano, G., and Ohtani, T.

(2003). A Gender and Age Estimation System from

Face Images. In SICE Annual Conference, volume 1,

pages 756–761. IEEE.

Jain, A. K., Dass, S. C., and Nandakumar, K. (2004). Soft

Biometric Traits for Personal Recognition Systems. In

Biometric Authentication, pages 731–738. Springer.

Kawano, T., Kato, K., and Yamamoto, K. (2005). An Anal-

ysis of the Gender and Age Differentiation using Fa-

cial Parts. In IEEE International Conference on Sys-

tems, Man and Cybernetics, volume 4, pages 3432–

3436. IEEE.

Kazemi, V. and Sullivan, J. (2014). One Millisecond Face

Alignment with an Ensemble of Regression Trees. In

IEEE Conference on Computer Vision and Pattern

Recognition, pages 1867–1874. IEEE Computer So-

ciety.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Im-

ageNet Classiﬁcation with Deep Convolutional Neu-

ral Networks. In Pereira, F., Burges, C. J. C., Bottou,

L., and Weinberger, K. Q., editors, Advances in Neu-

ral Information Processing Systems, volume 25, pages

1097–1105. Curran Associates, Inc.

Kwon, Y. H. and da Vitoria Lobo, N. (1994). Age Classiﬁ-

cation from Facial Images. In IEEE Computer Society

Conference on Computer Vision and Pattern Recogni-

tion, pages 762–767. IEEE.

Lanitis, A., Taylor, C. J., and Cootes, T. F. (2002). To-

ward Automatic Simulation of Aging Effects on Face

VISAPP 2018 - International Conference on Computer Vision Theory and Applications

400

Images. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 24(4):442–455.

Liu, K.-H., Yan, S., and Kuo, C.-C. J. (2015). Age Es-

timation via Grouping and Decision Fusion. IEEE

Transactions on Information Forensics and Security,

10(11):2408–2423.

Luu, K., Seshadri, K., Savvides, M., Bui, T. D., and Suen,

C. Y. (2011). Contourlet Appearance Model for Facial

Age Estimation. In International Joint Conference on

Biometrics, pages 1–8. IEEE.

Menotti, D., Chiachia, G., Pinto, A., Schwartz, W. R.,

Pedrini, H., Falc

ao, A. X., and Rocha, A. (2015).

Deep Representations for Iris, Face, and Fingerprint

Spooﬁng Detection. IEEE Transactions on Informa-

tion Forensics and Security, 10(4):864–879.

Murphy, K. P. (2012). Machine Learning: A Probabilistic

Perspective. MIT Press.

Ojala, T., Pietik

ainen, M., and M

aenp

a, T. (2002). Mul-

tiresolution Gray-Scale and Rotation Invariant Tex-

ture Classiﬁcation with Local Binary Patterns. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, 24(7):971–987.

Paulo Carlos, G., Pedrini, H., and Schwartz, W. R. (2015).

Classiﬁcation Schemes based on Partial Least Squares

for Face Identiﬁcation. Journal of Visual Communica-

tion and Image Representation, 32:170–179.

Pinto, A., Pedrini, H., Schwartz, W. R., and Rocha, A.

(2015). Face Spooﬁng Detection through Visual

Codebooks of Spectral Temporal Cubes. IEEE Trans-

actions on Image Processing, 24(12):4726–4740.

Ren, H. and Li, Z.-N. (2014). Age Estimation based on

Complexity-aware Features. In Asian Conference on

Computer Vision, pages 115–128. Springer.

Scikit-Learn Machine Learning in Python (2017).

http://scikit-learn.org/stable/.

Silva, P., Luz, E., Baeta, R., Pedrini, H., Falc

ao, A. X.,

and Menotti, D. (2015). An Approach to Iris Con-

tact Lens Detection based on Deep Image Represen-

tations. In 28th SIBGRAPI Conference on Graphics,

Patterns and Images, pages 157–164. IEEE.

Silva Pinto, A., Pedrini, H., Schwartz, W., and Rocha, A.

(2012). Video-based Face Spooﬁng Detection through

Visual Rhythm Analysis. In 25th SIBGRAPI Confer-

ence on Graphics, Patterns and Images, pages 221–

228. IEEE.

Suard, F., Rakotomamonjy, A., Bensrhair, A., and Broggi,

A. (2006). Pedestrian Detection using Infrared images

and Histograms of Oriented Gradients. In IEEE Intel-

ligent Vehicles Symposium, pages 206–212.

Suo, J., Zhu, S.-C., Shan, S., and Chen, X. (2010). A Com-

positional and Dynamic Model for Face Aging. IEEE

Transactions on Pattern Analysis and Machine Intel-

ligence, 32(3):385–401.

Thukral, P., Mitra, K., and Chellappa, R. (2012). A Hi-

erarchical Approach for Human Age Estimation. In

IEEE International Conference on Acoustics, Speech

and Signal Processing, pages 1529–1532. IEEE.

Weng, R., Lu, J., Yang, G., and Tan, Y.-P. (2013). Multi-

feature Ordinal Ranking for Facial Age Estimation.

In IEEE International Conference and Workshops on

Automatic Face and Gesture Recognition, pages 1–6.

IEEE.

Wu, J. and Rehg, J. M. (2011). CENTRIST: A Visual

Descriptor for Scene Categorization. IEEE Transac-

tions on Pattern Analysis and Machine Intelligence,

33(8):1489–1501.

Wu, T., Turaga, P., and Chellappa, R. (2012). Age Estima-

tion and Face Veriﬁcation across Aging using Land-

marks. IEEE Transactions on Information Forensics

and Security, 7(6):1780–1788.

Zhu, Q., Yeh, M.-C., Cheng, K.-T., and Avidan, S. (2006).

Fast Human Detection Using a Cascade of Histograms

of Oriented Gradients. In IEEE Computer Society

Conference on Computer Vision and Pattern Recog-

nition, volume 2, pages 1491–1498.

Combination of Texture and Geometric Features for Age Estimation in Face Images

401