Human Visual System Based Framework For Gender Recognition

Cherinet G. Zewdie

1,2

and Hubert Konik

1,2

Universit

e de Lyon, CNRS, 37 Rue du Repos, 69007 Lyon, France

Universit

e Jean Monnet, Laboratoire Hubert Curien, UMR5516, 42000 Saint-Etienne, France

Keywords:

Human Visual System, Gender Recognition, Salient Region, HVS Inspired Gender Recognition, Local Binary

Pattern.

Abstract:

A face reveals a great deal of information to a perceiver including gender. Humans use speciﬁc information

(cue) from a face to recognize gender. The focus of this paper is to ﬁnd out this cue when the Human Visual

System (HVS) decodes gender of a face. The result can be used by a Computer Vision community to develop

HVS inspired framework for gender recognition. We carried out a Pyscho-visual experiment to ﬁnd which

face region is most correlated with gender. Eye movements of 15 observers were recorded using an eye tracker

when they performed gender recognition task under controlled and free viewing condition. Analysis of the

eye movement shows that the eye region is the most correlated with gender recognition. We also proposed a

HVS inspired automatic gender recognition framework based on the Psycho-visual experiment. The proposed

framework is tested on FERET database and is shown to achieve a high recognition accuracy.

1 INTRODUCTION

A human face reveals a great deal of information to

a perceiver. These information include Gender, Eth-

nicity, Facial expression etc. Gender is an important

demographic attribute of people.

The Human Visual System (HVS) has an amazing

ability to decode gender of a face across different cul-

tures and ethnicities with few exceptions (Bruce et al.,

1993) in a very short time. Only speciﬁc face region

or cue is used by HVS to derive the gender informa-

tion from a face (Bruce et al., 1993)(Ng et al., 2012).

That is, only salient face region correlated with gen-

der recognition is used, where salient means the most

important.

The ability to recognize gender using computer vi-

sion has many applications in surveillance, human-

computer interaction, content-based indexing and

retrieval, biometrics, targeted advertizing etc. In

human-computer interaction, it helps to design hu-

manoid robots with knowledge to address humans ap-

propriately (as Mr or Ms). In content-based indexing

and retrieval, the ability to recognize gender helps in

annotating the gender of the people in an image for

the millions of images on the internet. In biometrics

application such as face recognition, the ability to rec-

ognize gender can cut the time required for searching

by half by training a separate classiﬁer for male and

female. In targeted advertizing, a computer vision

powered billboards can display advertisements rele-

vant to the person looking at the billboard based on

gender (Ng et al., 2012).

Different methods for automatic gender recog-

nition have been proposed (Alexandre, 2010)(Gutta

et al., 1998)(Andreu and Mollineda, 2008)(Kawano

et al., 2004)(Buchala et al., 2004)(Moghaddam and

Yang, 2002)(Makinen and Raisamo, 2008). However,

none of the proposed methods try to mimic the way

HVS performs gender recognition. These methods

rely on the power of image processing by spending

computational time on whole face image to extract ge-

ometric based information using measurements of fa-

cial landmark or appearance based information using

some operations or transformations applied on pixels

of the face image with the idea that such extracted in-

formation is different for female and male.

Luis A. Alexandre (Alexandre, 2010) proposed a

multi scale approach (considered as one of the state of

the art results) to solve the problem of gender recog-

nition. He proposed extracting features from different

face image resolutions and then a classiﬁer (SVM)

is trained for each feature type. The decision from

each classiﬁer is fused using majority vote to obtain

the ﬁnal classiﬁcation result. For each resolution, he

extracted intensity, shape and texture features. His-

tograms of edge directions were used as shape fea-

254

G. Zewdie C. and Konik H..

Human Visual System Based Framework For Gender Recognition.

DOI: 10.5220/0005176102540260

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence (ICAART-2015), pages 254-260

ISBN: 978-989-758-074-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

tures, pixel values as intensity and uniform LBP as

texture features. He tested this approach on a pre-

processed FERET database (geometric normalization

such as the same eye position, face rotation, and plac-

ing mouth in a ﬁxed position) and reported only one

miss classiﬁed image out of 107 test images (accu-

racy of 99.07 %). The high accuracy reported here is

as a result of a very complex approach that may not

be suitable for real time applications.

Other methods try to divide the face into regions

with the hypothesis that processing only such infor-

mation could discriminate between female and male.

Y.Andreu et al (Andreu and Mollineda, 2008) pro-

posed a method that exploits the role of facial parts in

gender recognition. Given the image of a face, a num-

ber of subimages containing mouth, eye, nose, chin;

and an internal face (made up of eye, nose, mouth

and chin), external face (made up of hair, ears and

contour) and the full face are extracted and converted

to appearance based data vectors. Classiﬁcation us-

ing SVM, KNN, and Quadratic Bayes Normal Clas-

siﬁer (QDC) tested on FERET (Phillips et al., 1998)

(2147 images) and XM2VTS (Andreu and Mollineda,

2008) (1378 images) showed that individual parts in-

clude enough information to be able to discriminate

between genders with accuracy of above 80% and the

joint contribution of the parts (internal face) produced

accuracy of over 95%. The motivation to evaluate

the contribution face parts was to deal with situations

where face images are partially occluded.

However, unlike these methods, we propose to

solve the problem of gender recognition by simulating

the HVS. We argue that the task of gender recognition

can be done in a more conducive manner, if only face

region correlated with gender recognition (salient re-

gion) is processed as it happens in HVS.

We conducted a Pyscho-visual experiment to ﬁnd

facial region (salient region) associated with gender

recognition. An eye-tracking device which records

ﬁxations and saccades has been used for our experi-

ment. Fixations are used to describe the visual atten-

tion when it is directed towards a salient region for the

assigned task and the eye gathers most of the infor-

mation during ﬁxations. Saccades are eye movements

between ﬁxations. We also propose a human visual

system based novel framework for gender recogni-

tion. The proposed framework creates a new fea-

ture space by extracting Uniform Local Binary Pat-

tern (ULBP) (Ojala et al., 2002) features from only

identiﬁed salient region of the face. The result from

our experiment can also be used by the computer vi-

sion community to develop robust algorithms by iden-

tifying salient region containing discriminative infor-

mation for gender recognition.

Our contribution in this study is three fold

1. Using a pyscho-visual experiment, we have statis-

tically and using gaze maps identiﬁed the face re-

gion (salient region) correlated with gender (both

male and female) according to human vision.

2. We used a novel approach of conducting an exper-

iment using eye tracker to ﬁnd gender cue from a

face and the reported result is validated in the do-

main of human behavior study and cognitive sci-

ence.

3. We proposed a novel HVS based framework

for automatic gender recognition which achieves

high recognition accuracy by only processing the

salient region.

The rest of the paper is organized as follows: in

section 2, we will present the psycho-visual exper-

iment we carried out to record eye movements (lo-

calize salient region). Section 3 discusses the result

of the analysis of the experiment. In section 4, we

will present the proposed automatic gender recogni-

tion framework. Results and Discussion of the auto-

matic recognition are presented in section 5. Conclu-

sion and Future works are given in section 6.

2 PYSCHO-VISUAL

EXPERIMENT

A gender recognition task is assigned to the observers.

We then recorded their eye movements under con-

trolled and free viewing condition. The recorded eye

movements is analyzed to ﬁnd which face region is

salient for the displayed gender (stimulus).

2.1 Participants and Stimuli

Fifteen students and teachers from the University of

Jean Monnet volunteered for the experiment. The

subjects were between 20 and 34 years of age. They

all had normal or corrected to normal vision. They

were given a short brieﬁng about the experiment and

the apparatus before the start of the experiment.

The stimuli were created by choosing 20 gray

scale face images (10 female and 10 male) from

FERET dataset (Phillips et al., 1998) with a criterion

of maximizing diversity. The criterion includes differ-

ent ethnicities, faces with special characteristics (such

as glass, beard), different ages (young and old).

2.2 Apparatus

A video based eye tracker called EyeLink II from SR

research is used to record the eye movements of the

HumanVisualSystemBasedFrameworkForGenderRecognition

255

human observers as they performed the gender cat-

egorization task. The EyeLink II is equipped with

three miniature infrared cameras with one mounted

on a lightweight headband for head motion compen-

sation and the other two mounted on arms attached to

headband for tracking both eyes.

The stimuli were presented on a 19 inch CRT

monitor with a resolution of 1024 x 768 and a refresh

rate of 85 Hz. The viewing distance was 70 cm result-

ing in a visual angle of 5.76

x 6.42

2.3 Procedure

We preformed the experiment in a dark and quiet

room to avoid any situation that would distract the ob-

servers. There was nothing in front of the observers

ﬁeld of view except the stimulus. The experiment was

designed in such a way that it won’t be tiresome for

the observers or loose interest in the experiment.

2.4 Eye Movement Recording

The eye position of the observers was tracked at 500

Hz with an average noise of 0.010. Before the record-

ing process started, we performed calibration and val-

idation. Calibration is used to collect ﬁxations on tar-

get points, in order to map raw eye data to gaze posi-

tion. A nine point calibration is chosen for this experi-

ment. For observers, the calibration step is ﬁxating on

a nine sequentially and randomly displayed points on

different locations of the screen. The validation step

is used to ﬁnd the gaze accuracy of the calibration. A

threshold error of 1

has been selected as the greatest

divergence that could be accepted.

The head mounted eye tracker compensates for

head movements so that observers have the ﬂexibil-

ity to perform experiments in a free viewing condi-

tion. The data collected from the experiment is also

not affected by the head movements as a result of this

compensation.

3 RESULTS AND DISCUSSIONS

The primary data that we obtain using the eye tracker

is the ﬁxations and saccades of observers collected

while performing the gender categorization task. Two

types of analysis can be done on the collected ﬁx-

ations once eye blinks are separated and removed.

They are gaze map construction and statistical anal-

ysis of the ﬁxations. However, since our stimuli are

static, it is very important that we know in advance the

duration of the ﬁxations that we analyze. Each of the

20 stimulus was presented to each observer for 3000

ms. But we analyzed the ﬁxations collected during

the ﬁrst 800 ms. This is because, for a deﬁned gender

recognition task, when an observer is presented with

a stimulus, the observer recognizes gender around

≈ 200 − 250ms after stimulus onset (Mouchetant-

Rostaing et al., 2000). However, the categorization

task sometimes happen around ≈ 250 − 700ms. To

include all possibilities, ﬁxations during the ﬁrst 800

ms were analyzed. We have also removed the ﬁrst ﬁx-

ation as this is related to the cross sign that precedes

every stimulus presentation so that all observers start

the task from the same center face location.

Figure 1: Gaze maps for gender recognition.

3.1 Gaze Map Construction

The simple, however, important output that we can

obtain from the recorded ﬁxation points is a gaze map.

Figure 1 gives the ﬁrst impression that gazes from

all observers for both female and male stimuli are at-

tracted to eyes, nose and mouth face regions. The sec-

ond important fact we draw from the gaze map is that

the eyes region is the most salient. This is because

the superimposed color blobs is warmer for the eyes

region compared to the mouth and the nose regions.

That means, most of the ﬁxations from the ﬁfteen ob-

servers were attracted to the eyes region.

Figure 2: Saccades analysis for a face with special charac-

teristic.

3.2 Statistical Analysis of Fixations

The conclusion that we made from the gaze map anal-

ysis is that the eyes region is the most salient for gen-

der recognition task. We need to conﬁrm this conclu-

sion through statistical analysis of the collected ﬁxa-

tions for different face regions. To conﬁrm this, we

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

256

computed and statistically analyzed the average per-

centage of trial time observers have ﬁxated in the dif-

ferent face regions (eyes, nose and mouth) during the

ﬁrst 800ms for each stimulus presentation.

Figure 3: Statistical analysis of ﬁxations for both female

and male stimuli.

Figure 3 conﬁrms the conclusion made from the

gaze map shown in ﬁgure 1 about the saliency of eye

region for gender recognition task. That is observers

use most of the information from the eye region to

categorize a stimulus as female or male. It can be eas-

ily observed from ﬁgure 3 that the average percentage

of trial time spent is around 54.3%, 25.4% and 2.8%

in the eyes, nose and mouth regions, respectively for

male face; and 58.6%, 26.3% and 3.9%, respectively

in the eyes, nose and mouth regions for female face.

For both genders, the amount of time spent in the eye

region is more than twice the amount spent in the

nose region. The error bar represents the standard

error (SE) of the mean. The saliency of the eye re-

gion is further conﬁrmed by the analysis of saccades

shown in ﬁgure 2. In the ﬁgure, we see a face stimulus

with a beard. Beard is a characteristic associated with

male face. Even in the presence of this characteris-

tic, observers have still searched for gender informa-

tion from the eye region. All the eye movements be-

tween ﬁxation (saccades) were in the eye region. The

discovery of the eye region as carrying information

for gender recognition is consistent with the result by

Brown et al (BrownU and Perrett, 1993) where six-

teen male faces were averaged to create a male proto-

type and sixteen female faces where averaged to cre-

ate a female prototype. Individual face regions such

as eyes, brows, nose, mouth and chin were shown to

subjects for gender recognition. The authors reported

that all regions except the nose carried some informa-

tion about gender.

4 PROPOSED AUTOMATIC

GENDER RECOGNITION

FRAMEWORK

Using the conclusion from section 3, we are going to

test the idea that algorithmically gender can be recog-

nized by processing only salient region. The proposed

Figure 4: Human Visual System based feature extraction

framework from a face for gender recognition

framework is shown in ﬁgure 4.

4.1 Localize Salient Region

As shown in ﬁgure 4, for a given face, we ﬁrst de-

tect the interest region. That means we try to detect

the eye region since the eye region is the most salient

for gender recognition. There are different detection

algorithms. We chose to use the detection algorithm

proposed by Viola and Jones (Viola and Jones, 2004)

since the algorithm has a minimum false positives

compared to other detection algorithms. Once we de-

tect the eye region, we localize that region as the re-

gion from where we extract the features.

4.2 Feature Extraction

Once the eye region is localized, the next task would

be to represent it using feature vectors(descriptors).

There are different feature extraction methods. How-

ever, the type of features they extract (intensity, shape,

texture etc) and their discriminative power is differ-

ent depending on the problem at hand. The Uniform

Local Binary Pattern (ULBP), which is the extension

of the Local Binary Patterns (Ojala et al., 2002) is

shown in different studies to have a high discrimina-

tive power for gender recognition (Ng et al., 2012).

Based on their track record, we used ULBP to extract

texture features from the salient eye region.

When describing a face with LBP patterns, it is

important to consider local features instead of global

features. For this reason, we computed the ULBP us-

ing a small local window and concatenated the fea-

tures together. The other reason for doing this is that

since the salient region we are considering is only

the eyes region and considering the fact that there are

similarities between female face eye region and male

face eye region, the discriminative power of features

could be improved if we capture the information lo-

HumanVisualSystemBasedFrameworkForGenderRecognition

257

cally. This approach is used by different researchers

(Alexandre, 2010) even if the features were not ex-

tracted from a salient region but the whole face re-

gion.

4.3 Classiﬁers

Once the images are represented with features with a

high degree of discriminative power, the next step will

be to feed these feature vectors to a machine learning

algorithm that learns a model for each gender. Gender

recognition is a binary classiﬁcation. A face image

can only be a male or female. So one class is posi-

tive and the other is negative. If a given face image is

not classiﬁed as a male (which in our study is a posi-

tive class) then it will automatically be classiﬁed as a

female. There are a number of machine learning al-

gorithms to solve the gender recognition problem. We

tested SVM, KNN, C4.5 Decision tree and ensemble

methods such as Bagging, Random Forest and Ad-

aboost.

5 RESULTS FOR AUTOMATIC

GENDER RECOGNITION

5.1 Database Selection

The challenging problem when solving gender recog-

nition is the selection of database to test on since

there is no speciﬁc database (containing multiple eth-

nicities) made for gender recognition. Limited by

this problem, just like other researchers, we also

selected images from a facial recognition database

called FERET (Phillips et al., 1998) to test our algo-

rithm. In total, we selected 207 images from FERET

database to test our algorithm. 127 of the total is used

for training and 80 is used for testing. The distribution

for male and female is shown in table 1. No images

of the same person is repeated in the database.

Table 1: Distribution of images as training and test set se-

lected from FERET database.

Original size Total number Training Test

Female male Female Male

FERET 512X768 207 60 67 40 40

5.2 Results

The classiﬁers listed above were used for learning and

classifying the face images as either gender. The re-

sults from these classiﬁers are shown in table 2. The

displayed accuracy value in the table is the accuracy

of each classiﬁer on the unseen data (on the 80 test

images). A 5-fold cross validation is used to choose

parameters for the classiﬁers and kernels used. In ad-

dition to measuring the accuracy in terms of correctly

classiﬁed test images, we have also used ROC curve

(value) to evaluate the performance of our algorithm.

Table 2: Classiﬁcation results of the proposed framework

for different classiﬁers.

ACCURACY % Alexander et al (Alexandre, 2010) State of the art

SVM + RBF +

Cross validation (Weka)

100 99.01

KNN +Cross validation 100 -

C4.5 decision tree 100 -

Bagging 100 -

Random Forest 100 -

Adaboost (Tree) 100 -

The result of the analysis of ROC curve for SVM,

KNN, C4.5 Decision tree, Bagging, Random Forest

and Adaboost is shown in ﬁgure 5.

Figure 5: Receiver Operating Characteristics (ROC) for

SVM, C4.5, KNN, Bagging, Random Forest, Adaboost.

All the accuracy results presented in table 2 were

obtained on the 80 test images. Each of the classiﬁers

shown in the table classiﬁed the given test images into

their respective classes. We believe that different fac-

tors have contributed to the reported high accuracy

of the result. The ﬁrst reason is the processing of

only salient region of the face correlated with gender

recognition. This is because features with no discrim-

inative power complicates the model that the classi-

ﬁer learns and thus reduce the predicting capability of

the learned model. However, exploring the HVS to

ﬁnd the face information used by human when rec-

ognizing gender and using only those discriminative

information with computer vision algorithms to rec-

ognize gender would produce an impressive result as

shown in the table. The second reason is the type of

features (texture) used and how the features are ob-

tained. Texture features are shown in different studies

to have produced a good result for gender recognition

problem. Even if we claim that the salient region we

used has a discriminative power, it may not produce

the result shown in the table if the information from

this region is not captured properly. Since the eye

salient region for both male and female bears resem-

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

258

blance, capturing the features at a global level doesn’t

help in discriminating between the eye region of male

face and female face. For this reason, we captured the

information at a local level by using a small window.

This helps us to capture the small spatial information

of the eye region for both gender. The window covers

the full image size by sliding horizontally and verti-

cally. The result of the other window sizes we tried

is shown in table 3. The third reason is the choice of

classiﬁers. SVM, KNN and C4.5 are shown to pro-

duce a good result for a gender recognition problem

in the survey made by (Ng et al., 2012). It is also no

surprise, in addition to the quality features we had,

that we obtained excellent result with Bagging, Ran-

dom Forest and Adaboost since they combine results

from a number of week classiﬁers. The number of

weak classiﬁers (trees) used are 800, 500 and 90 for

Bagging, Radom Forest and Adaboost, respectively.

Table 3: Accuracy(%) of different window size for different

classiﬁers.

Window size Bagging Random Forest Adaboost KNN SVM

6x5 100 100 100 100 100

10x10 73.75 71.25 75 63.75 73.75

10x12 73.75 73.75 78.75 60 75.25

15x20 63.75 67.5 72.5 56.25 68.5

30x60 65 66.25 62.5 65 59.25

We have also compared our result with the method

considered as one of the state of the art results pro-

posed at (Alexandre, 2010). As we repeatedly ex-

plained before, it is difﬁcult to compare results from

gender recognition algorithms since authors use dif-

ferent databases to test their algorithm. However,

comparisons are usually made in the literature be-

tween algorithms tested on images from the same

database (not necessarily the same images). For this,

since we selected the images from a FERET (Phillips

et al., 1998) dataset, we compared our result with

(Alexandre, 2010) since this method is also tested on

FERET as shown in table 2. One miss classiﬁed im-

age is reported in (Alexandre, 2010) from 107 images.

However, the high recognition accuracy is obtained

owing to the very complex approach used. Consid-

ering the complexity of the proposed multiscale deci-

sion fusion approach, we believe that our method is

simple and achieves a very good accuracy.

6 CONCLUSION AND FUTURE

WORK

This study presented the face region correlated with

gender recognition for both female and male faces by

studying the HVS. This conclusion is made by record-

ing eye movements of 15 observers as they performed

a gender recognition task on 20 images chosen from

FERET database. The constructed gaze map and sta-

tistical analysis of the collected ﬁxations show that the

eye region is the most salient for gender recognition.

The localized salient region can be used by the com-

puter vision community to have an insight from where

to extract discriminative descriptors as this is the most

important step when developing robust and efﬁcient

automatic gender recognition algorithms. In addition

the extraction of descriptors from only the salient re-

gion would result in a fast system as it reduces the

computational complexity of the algorithms.

We also proposed a novel framework for au-

tomatic gender recognition based on the localized

salient region. We have achieved a high recognition

accuracy by processing only salient region of a face.

In this paper, we only considered frontal faces

when we created stimuli for the experiment. In the

future, we have plans to include different face orien-

tations into the stimuli and see if the same conclusion

can be made. We also intend to occlude the eye region

and ﬁnd out if secondary information can be used for

gender recognition, (i.e if gender recognition is hier-

archical). Including different image resolution in the

stimuli is also one of the many works we planned in

the future.

ACKNOWLEDGEMENTS

This paper is supported by the Erasmus Mundus

Scholarship 2012-2014 sponsored by the European

Union, and CNRS.

REFERENCES

Alexandre, L. A. (2010). Gender recognition: A multiscale

decision fusion approach. Pattern Recognition Let-

ters, 31(11):1422–1427.

Andreu, Y. and Mollineda, R. A. (2008). The role of face

parts in gender recognition. In Image Analysis and

Recognition, pages 945–954. Springer.

BrownU, E. and Perrett, D. (1993). What gives a face its

gender. Perception, 22:829–840.

Bruce, V., Burton, A. M., Hanna, E., Healey, P., Mason, O.,

Coombes, A., Fright, R., and Linney, A. (1993). Sex

discrimination: how do we tell the difference between

male and female faces? Perception.

Buchala, S., Davey, N., Frank, R. J., Gale, T. M., Loomes,

M. J., and Kanargard, W. (2004). Gender classiﬁca-

tion of face images: The role of global and feature-

based information. In Neural Information Processing,

pages 763–768. Springer.

HumanVisualSystemBasedFrameworkForGenderRecognition

259

Gutta, S., Wechsler, H., and Phillips, P. J. (1998). Gen-

der and ethnic classiﬁcation of face images. In Auto-

matic Face and Gesture Recognition, 1998. Proceed-

ings. Third IEEE International Conference on, pages

194–199. IEEE.

Kawano, T., Kato, K., and Yamamoto, K. (2004). A com-

parison of the gender differentiation capability be-

tween facial parts. In Pattern Recognition, 2004. ICPR

2004. Proceedings of the 17th International Confer-

ence on, volume 1, pages 350–353. IEEE.

Makinen, E. and Raisamo, R. (2008). Evaluation of gen-

der classiﬁcation methods with automatically detected

and aligned faces. Pattern Analysis and Machine In-

telligence, IEEE Transactions on, 30(3):541–547.

Moghaddam, B. and Yang, M.-H. (2002). Learning gen-

der with support faces. Pattern Analysis and Machine

Intelligence, IEEE Transactions on, 24(5):707–711.

Mouchetant-Rostaing, Y., Giard, M.-H., Bentin, S., Aguera,

P.-E., and Pernier, J. (2000). Neurophysiological cor-

relates of face gender processing in humans. Euro-

pean Journal of Neuroscience, 12(1):303–310.

Ng, C. B., Tay, Y. H., and Goi, B. M. (2012). Vision-based

human gender recognition: a survey. arXiv preprint

arXiv:1204.1611.

Ojala, T., Pietikainen, M., and Maenpaa, T. (2002). Mul-

tiresolution gray-scale and rotation invariant texture

classiﬁcation with local binary patterns. Pattern Anal-

ysis and Machine Intelligence, IEEE Transactions on,

24(7):971–987.

Phillips, P. J., Wechsler, H., Huang, J., and Rauss, P. J.

(1998). The feret database and evaluation procedure

for face-recognition algorithms. Image and vision

computing, 16(5):295–306.

Viola, P. and Jones, M. J. (2004). Robust real-time face

detection. International journal of computer vision,

57(2):137–154.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

260