Kiểm soát truy cập bằng sử dụng nhận dạng khuôn mặt

Viet Nam Received' December 10, 2013; accepted: April 22, 2014 Abstract In this paper, we propose an access control system using human face recognition Basically, the system composes o

Access Control Using Face Recognition

Pham Thi Thanh Thuy\ Le Thi Lan, Dao Trung Kien, Pham Ngoc Yen

Hanoi University ofScience and Technology

No 1 Dai Co Viet Str, Ha Noi Viet Nam Received' December 10, 2013; accepted: April 22, 2014

Abstract

In this paper, we propose an access control system using human face recognition Basically, the system composes of two mam parts- face recognition and door control The contnbution of our paper is the Pattern Histogram) feature matching and classification of minimum distance decision mle The advantages

of LBPH features Is that they have high speed of matching and suitable for real-time face recognition applications We have compared the performance of the face recognition method using LBPH with two state of the art methods based on LDA (Linear Discriminant Analysis) and PCA (Principal Component Analysis) The expenmental results show that the proposed method outperforms the state of the art methods Based on this result, m order to open door automatically, we design a hardware circuit which door to be opened or still closed based on the face recognition result

Keywords Face recognition, LBPH features, access control

1 Introduction

Biometric identification is applied m man\

securit\ surveillance and access control systems In

companson with other biometnc features such as ms

sources and requires less controlled mteraution

Human face recognition has attracted many

researchers over decades howe\er this is still -i

feature extraction Optimal selected and high

discnrm native features result to good object

recognition with low computational cost and

beneficial to the real hme processing svstem Two

approaches are proposed for finding optimal face

features, those are local and global teature extraction

The last one, such as PCA or LDA, focus on the

whole face image, otherwise the first one, for

example LBP or Gabor features [1,2], uses some

unique identifying features on human face As a

features are normally not as good as local features

[3]

The contnbution of our paper is the automatic

access control system which uses face recognition

method based on LBPH feature matching and

classification of minimum distance decision rule The

advantages of LBPH features is that they have high

' Conespongding Author: Tel, (+84) 915 651.748

speed of matching and suitable for real-time face recognition applications Moreover, they are less sensitive to illumination variaUons [4], LBPH features are extracted in the detected face patches that are processed in this paper help build high distinguishable feature vectors and give better result circuit which connects the computer with an electrical door This circuit has the funchon of a normal lock When a "legal key" of the owner's face

IS available, it will unlock door automatically

2 Automatic access control using human face recognition

The system includes an IP camera which captures video streams, a hardware circuit which is connected to a computer through COM port, a computer which acquires and processes video streams, recognizes human faces and controls hardware circuitto unlock door (Fig.l)

[^^M

Fig 1 Access control system using face recognition Face recognition and door opening control modules will be explained in detail in the following sections

2.1 Face recognition module

Face recognition module basically includes

five steps of processing' input image acquisition, face

detection, face image preprocessing, face feature

extraction and recognition (Fig, 2) The last one

covers two phases: training (learning) and evaluating

(testing) In the training phase, the system will be

learned from face image database, then individual

facial model will be established In the evaluating

phase, the system will give the result of

identification/verification for the testing facial image

based on the trained models

Input image acquisition: Video streams are captured

from IP camera and each frame will be processed for

face detection

Face detection: We use Haar like features and

Adaboost cascade classifier algonthm for face

detection [5] The results of face detection are face

patches (region bounded by red rectangle in (Fig 3),

affect the later face recognition process, a face patch

only includes the region of eyebrows, eyes, nose and

mouth

Face Image Preprocessing: The preprocessing

techniques are applied for fece patches so some major

challenges for face recognition can be solved First,

the original face images have to be converted to the

grayscale form Then, histogram equalization

techmque is applied to tackle illumination vanations

[6], Face patches are also aligned to adjust orientation

and resized to 100x100 pixels

•rizrH

1

1

\

RecogniliDD

Tnucmg

Evaluating

[ Image Preprocessing

r - j Feature

/

Fig 2 Human face recognition system

Fig 3 Face defection result (region bounded by red

rectangle)

Feature Extraction: It is not effective if we use face

patches directly for face recognition The

these face patches are taken from different camera

extraction provides dimension reduction, salient feature extraction and noise removal After this process, each face patch is converted to a vector with fixed dimension

In this paper LBPH features are used for face description Instead of using entire face patch, we extract LBP local features from this The basic idea

of LBP method is to summanze the local structure in

a face patch by comparing each pixel with its neighbors Taking a pixel as center, if the intensity of center pixel is greater-equal its neighbor, then denote

it with 0 and 1 if not (1) The result is a binary sequence Using 8 surrounding pixels we will end up with 28 possible combinations, called Local Bmary Pattems (Fig 4)

i f i ^ P , - I ^ ( g - g c ) 2 ' \, if T>0;

0, otherwis

(1)

Where LBPRR is LBP code with P sampling points

on a circle of radius of R, gc is the grey value of the

pixels

LBP local feamre s are implicit low-dimensional and simple in computation Thus, it is possible to analyze face image in challenging real-time settings Besides, they are more robust against

methods such as PCA, LDA [8]

Fig 4 An example of LBP computation [7]

Fig 5 LBP images with different gray-scale transformations

oCfKcimagt ffomeachblock

Fig 6 Face description with LBPH

Fig 7 Access control

After calculation of LBP face patch, we divide

il to 64 sub regions (8x8) and extract histogram from

each (Fig 6) A feature vector is then obtained by

concatenating the local histograms These histograms

are called Local Binary Pattems Histograms (2)

^, = E ^ f.i^>y) = 1 = 0,, (2)

Where f(x,y) is LBP of face patch or LBP of labeled

by the LBP operator; l[A] is I if A is true and 0 if A

is false

This histogram will be normalized to get a

coherent description for face classification

(3)

' E:X

Recognition: Training data of different preprocessed

face patches is created and stored in a folder It has a

single XML file that contains tags for the name of

person and a file name for the traming image Each

file identifiers can be generated This prevents images

being overwritten and easily allows several images

for one individual to be acquired and stored with no

problems

In evaluating process, testing images (face

patches that are taken from video frames) firstly

preprocessed, then their LBPH features are extracted

prepared traming data set and the results for face

decision rule

2.2 Door access control module

The result from face recognition is passed to door access connol module This module lake two tasks- input validatins and door open controlling (Fig.7)

The biometric face recognition includes two parts: identification (recognition) and verification (authentication) Given the testing face image, the first one points out the answer for the question "Who

is he''" That means the system will look in the traming database for nominated face which is the most similar with the probe face, for example he is true or not (Is it subject A'') This is done by threshold validation, A threshold is established by vector is higher than threshold, the verification result

one or a person who is contained m trainmg database

If the ID of the authenticated person is identical to the room's owner ID (validating process), this module

door automatically (door open controlling), otherwise

It is still closed

In order to open door automatically using the room owner's face, we set a Mmer of 2 seconds This parameter can be sel based on the application requirement During this time, if the face recogmhon accuracy is above 90%, the door access control module will active the hardware circuit lo unlock the door,

3 Experimental results

3.1 Experimental Environment

We evaluate the sysiem in our showroom (Fig 8-a) In this room, we set a smart room with an electrical door and a camera IP (Axis Ml054) which

is attached on one side of the door (Fig 8-b)

Fig 8 (a) Layout of showroom with a smart r

For the face recognition, we take two phases:

training and testing The n-aining face database is

acquired by the above mentioned camera All

subjects were requested to stand at a predefined

distance from the camera system (aboul lm)

In the testing phase, we test the sysiem with

two types of dataset- probe set and gallery set The

gallery set contains images of subjects participated to

training images acquisition process while probe set

training images acquisition process but also new

subjects The testing subjects are required lo stand

opposite to the camera aboul one meter and look

straight into the camera

3.2 Measurements

In order lo evaluate the performance of the

system, we use two cnteria as follows

TP + FP

TP

TP + FN

Fig 10 Ten images of one subject in traming

Table I The result comparison of three different methods (a) face recognition for gallery set; (b)face recognition for probe set; (c) computational lime for face recognition phase

Where TP (True Positive) is the number of conectly

wrongly recognized faces and FN (False Negative) is

the number of lost recognized faces,

3.3 Results

a Face recognition results

For face recognition evaluation, we prepare a

training data set of 200 face images of 20 subjects, 10

images for each subject with different head poses and

face expressions (Fig, 9, Fig 10) 20 subjects for

gallery set and 25 (including 5 strange subjects) for

probe set are used in the testing

The face recognition results are showTi in

Table 1 for gallery set and probe set The number of

testing frames for each subject is 300,

Companng with two other state of the art

methods PCA and LDA, the LBPH has the highest

precision and lowest time consuming for face

recognition processing (using the compuler of Intel®

core(TM) 2 duo, CPU T5870(g2.00GHz, RAM

2GB)

FAR Sensibility

1 LBPH 1

9 83

1 91.06 IDA 11.49

89 67

11! 52 "*" 1

1 86.6'!

LBPH LDA PCA

(b)

Procassing T i m e

I M 1 i i t e c o n d s / F r a m a l 12.5 15.7 18.7

b Access control results

The performance of the door access control is evaluated in the time interval of Al=2s, Dunng this time, about 10 video frames are processed for face recognition In each period of two seconds, if the

of above 90% (equivalent to more than 9 video frames with human faces recognized correctly), the door will be opened automatically, otherwise it remains closed Because during input image acquisition, there are some first frames have bail

objects have unready postures when they come to the camera, so we inihahze a count variable In the interval of At, with each time of correct recognition for the room o-wner, this variable will be increased by one, otherwise it will be reduced by one The door

to 9 in this interval or it is still closed As a result, for

25 testing subjects, the door is opened for 20 subjects

in gallery sel and closed for 5 unknown subjects

4 Conclusions

The result of experimentation shows promising results obtained fi'om our sysiem, However, these are only evaluated in confroiled environment In the fiiture work, we would like to develop an effective human identification model in

using others information source for example the

depth information from Kinect device or by

combining with other human identification method

door opening control should be developed into a

complete product like an automatic lock

Acknowledgments

The research leading to this paper was

supported by the National Project B2013.01.41

"Study and develop an abnormal event recognition

system based on compuler vision techniques"

References

[1] Ojala, T, M Pietikainen, and M Maenpaa,

Multi-resolution gray-scale and rotation invariant texmre

classification width local bmary pattems IEEE

Transactions on Pattern Analysis and Machine

Intelligence, 24 (2002) 971-^987,

[2] Ll, v., Z Ou, and G Wang, Face Recognition Using

Gabor Features and Support Vector Machines, in

and Y, Ong, Editors, Spnnger Berlin Heidelberg

(2005)119-122

[3] Da-Rui, S and W Le-nan A local-lo-hoIisUc face

recognition approach using elastic graph matching

Machine Learning and Cybernetics, 2002, Proceedmgs

2002 International Conference on, 2002,

[4] Al-Sahaf, H„ Z Mengjie, and M, Johnston Binary

image classification using genetic progranuning based

New Zealand (IVI^Z), 2013 28lh Intemalionai

Conference of

2013-[5] Viola, P and M, Jones, Robust Real-time Object

Detection, Second Intemalionai workshop on statistical

and computational theories of vision-modeling,

learning compuHng and sampling 2001 Canada

[6] ACHARYA T and A K RAY Image

Processing-Pnnciples and Applications 2005: Wiley InlerScience

[7] Dl, H,, el al, 1-ocai Binary Patterns and Its Application

to Facial Image Analysis A Survey, Systems, Man, and

Cybernetics, Part C Applications and Reviews, IEEE

Transactions on, 41 (2011)765-781,

[8] Ahonen, T„ A Hadid, and M, Pietikainen, Face

Recognition with Local Binary Patlems, ui Computer

Springer Berhn Heidelberg (2004)469^81