Leveraging Gabor Phase for Face Identiﬁcation in Controlled Scenarios

Yang Zhong and Haibo Li

Department of Computing Science and Communication, KTH Royal Institute of Technology, 100 44, Stockholm, Sweden

Keywords:

Face Recognition, Controlled Scenario, HD Gabor Phase, Block Matching, Learning-free, Deep Learning.

Abstract:

Gabor features have been widely employed in solving face recognition problems in controlled scenarios. To

construct discriminative face features from the complex Gabor space, the amplitude information is commonly

preferred, while the other one — the phase — is not well utilized due to its spatial shift sensitivity. In this

paper, we address the problem of face recognition in controlled scenarios. Our focus is on the selection of

a suitable signal representation and the development of a better strategy for face feature construction. We

demonstrate that through our Block Matching scheme Gabor phase information is powerful enough to im-

prove the performance of face identiﬁcation. Compared to state of the art Gabor ﬁltering based approaches,

the proposed algorithm features much lower algorithmic complexity. This is mainly due to our Block Match-

ing enables the employment of high deﬁnition Gabor phase. Thus, a single-scale Gabor frequency band is

sufﬁcient for discrimination. Furthermore, learning process is not involved in the facial feature construction,

which avoids the risk of building a database-dependent algorithm. Benchmark evaluations show that the pro-

posed learning-free algorithm outperforms state-of-the-art Gabor approaches and is even comparable to Deep

Learning solutions.

1 INTRODUCTION

Face recognition (FR) is a well established research

area and it has been studied for more than two decades

(Turk and Pentland, 1991; Belhumeur et al., 1997;

Wiskott et al., 1997; Ahonen et al., 2004; Zou et al.,

2007; Wright et al., 2009; Chan et al., 2013; Taigman

et al., 2013). Typically, face recognition works in two

essentially different modes: face veriﬁcation or face

identiﬁcation under either in controlled scenarios or

in the wild. Face veriﬁcation performs 1:1 match-

ing and provides a binary decision to the claimed

identity. Face veriﬁcation in controlled scenarios has

reached a rather high accuracy (Givens et al., 2013).

To tackle face veriﬁcation in uncontrolled scenarios,

many approaches have been proposed for more ef-

fective alignment (Cao et al., 2014; Yi et al., 2013;

Chen et al., 2012), utilization of different types of fea-

ture representations (Lowe, 2004; Dalal and Triggs,

2005; Ahonen et al., 2004) and matching metric for

comparing faces (Hua and Akbarzadeh, 2009; Pinto

et al., 2009; Li et al., 2013). Driven by innovation

in Deep Learning approaches, face veriﬁcation per-

formance has been greatly advanced in recent years.

By learning from big data, end-to-end artiﬁcial net-

works can outperform human on challenging veriﬁca-

tion tasks, e.g., the Labeled Face in the Wild (Huang

et al., 2007), (Huang et al., 2012; Taigman et al.,

2013; Sun et al., 2013; Schroff et al., 2015).

In contrast, face identiﬁcation is more difﬁcult. It

performs 1:N matching to sort out the gallery images

based on pair wise similarity measurements. Obvi-

ously, the operating requirement of face identiﬁca-

tion is vastly more demanding than operating merely

in veriﬁcation: an identiﬁer needs to be roughly N

times better than a veriﬁer to achieve comparable

odds against making false matches (Daugman, 2006).

This is probably why progress in face identiﬁcation

has been relatively insigniﬁcant over the last ﬁve

years. Though the proposed face identiﬁcation ap-

proaches have become increasingly complex, recog-

nition performance according to the benchmark eval-

uations remained relatively constant (Xie et al., 2010;

Yang et al., 2013; Cament et al., 2014; Chai et al.,

2014). To make a breakthrough in face identiﬁcation,

it seems we must revisit the foundation of face recog-

nition, and have a fresh look at the fundamental build-

ing blocks of face recognition.

A most fundamental building block of face recog-

nition is construction of features for measuring sim-

ilarity between two face images. The construction

of features consists of two steps: (1) the selection

of a suitable face representation; (2) feature extrac-

tion from the representation. There is a large collec-

Zhong, Y. and Li, H.

Leveraging Gabor Phase for Face Identiﬁcation in Controlled Scenarios.

DOI: 10.5220/0005723700490058

In Proceedings of the 11th Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2016) - Volume 4: VISAPP, pages 49-58

ISBN: 978-989-758-175-5

tion of research papers on how to extract stable, local

or global discriminative features, e.g., the commonly

used SIFT (Lowe, 2004), HOG (Dalal and Triggs,

2005), and LBP (Ahonen et al., 2004). Recently, these

features have been criticized as “hand-crafted”. It has

been claimed that better features can be learned auto-

matically from big face data collections through Deep

Learning approaches as in (Huang et al., 2012; Taig-

man et al., 2013; Sun et al., 2013).

While Deep Learning solutions generally have

signiﬁcant performance advantages over conventional

approaches, the dependence on access to very large

training datasets, the knowledge of designing deli-

cate hand-crafted artiﬁcial neural network and careful

engineering is considerable. This indicates that the

learned network could bias (or likely overﬁt) to the

training data; to adapt well to novel tasks, sufﬁcient

data for ﬁne tuning or even the construction of addi-

tional components in the learned architecture is often

a must (Parkhi et al., 2015) (we demonstrate this in the

last section). It also worth nothing that Deep Learn-

ing solutions are generally computational demanding:

typically the Convolutional Neural Networks (CNN)

involves millions of parameters. Thus, although we

believe a well tuned deep neural network is a most

effective FR solution (e.g., for FR in the wild), in

many scenarios where users are cooperative and envi-

ronments are controlled, it is often worth to consider

efﬁcient and effective solutions depending on much

shallower networks. Considering this, it is tenable to

revisit one of the most widely adopted approaches,

Gabor transformation, which paved the way for face

representation in controlled scenarios.

The Gabor transformation enables the employ-

ment of rich low-level multi-scale features by trans-

forming images from pixel domain to the complex

Gabor space. In the complex Gabor transformed

space, one reasonable option for many state-of-the-art

approaches was to utilize the amplitude for face rep-

resentation and feature construction. This is because

the amplitude varies slowly with spatial shift, mak-

ing it robust to texture variations caused by dynamic

expressions and imprecise alignment. By construct-

ing LBP-type features mostly from the amplitude and

applying various learning techniques, many Gabor

based approaches have shown remarkable advantages

over pixel-featured based methods: the identiﬁcation

rate in benchmark evaluations has benn found to be

improved by more than 20% (reaching around 90%)

thanks to the “blessing of dimensionality” (Givens

et al., 2013) ( but at the high cost of computational

efﬁciency (Mu et al., 2011; Chai et al., 2014)). Now

the question is how to achieve face identiﬁcation rates

in the range from 90% to 95% or even higher. In this

paper we argue that leveraging Gabor phase could en-

able such performance improvement for face identiﬁ-

cation.

The Gabor phase is robust to light change and

indeed has been well-known that phase is more im-

portant than amplitude for signal reconstruction (Op-

penheim and Lim, 1981). Gabor phase should have

played a more important role in face identiﬁcation.

However, use of Gabor phase in face recognition is

far from common and it has often been unsuccess-

ful with worse or nearly the same performance as

the amplitude in comparative experiments (Gao et al.,

2006; Zhang et al., 2009; Xie et al., 2010; Cament

et al., 2014). This is largely due to two challenging

issues: (1) Gabor phase is a periodic function and

a hard quantization occurs for every period; (2) it is

very sensitive to spatial shift (Wiskott et al., 1997;

Zhang et al., 2009), which imposes a rigid require-

ment on face image alignment. The ﬁrst issue was

partly solved by introducing the phase-quadrant de-

modulation technique (Daugman, 2004), but the sec-

ond issue is still far from being solved. The state-

of-the-art Gabor phase approach (LGXP (Xie et al.,

2010)) extracts varied LBP from the phase spectrum.

Since the combination of the phase and LBP is also

sensitive to spatial shift, the power of Gabor phase

was not demonstrated in face identiﬁcation.

In this work, we propose a method that merely

leverages the power of Gabor phase to address the

problem of face identiﬁcation in controlled scenarios.

We apply a slim ﬁlter bank of only two Gabor ﬁlters

to extract the Gabor phase information and perform

explicit matching on the quantized phase map via our

Block Matching scheme (Zhong and Li, 2014). Dif-

ferent from other elastic matching schemes, the Block

Matching scheme not only cancels the patch-wise

spatial shift in phase map but also simultaneously

evaluates the patch-wise utility during the learning-

free matching process. Combining the matching

scheme with phase codewords enables the exploit-

ment of high-diﬁnition phase information (4 times

higher than (Xie et al., 2010)) from only 2 Gabor ﬁl-

ters. Thus, the proposed approach can signiﬁcantly

bring up the algorithmic efﬁciency without sacriﬁcing

the recognition accuracy. Further more, it is totally

comparable to those state of the art Gabor solutions

and even CNN based solutions.

The disposition of our paper is as follows: we

ﬁrst brieﬂy review the related Gabor based and CNN

based solutions in Section 2; our approach is then de-

scribed in Section 3 followed by comparative experi-

ments presented in Section 4. We further compare the

performance between our approach and CNN archi-

tectures in Section 5, where we discuss our work as a

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

whole and offer our conclusions.

2 RELATED WORK

In this section, we ﬁrst describe the Gabor representa-

tion and then review recent Gabor based face recogni-

tion methods utilizing the Gabor amplitude or phase

in different ways.

In state-of-the-art face identiﬁcation methods

based on Gabor wavelet representations, there are two

steps to construct features for identiﬁcation. The ﬁrst

step is to use Gabor ﬁlters for an optimal image repre-

sentation. A Gabor face is obtained by ﬁltering a face

image with the Gabor ﬁlter function, which is deﬁned

as:

u,v

(z) =

u,v

(−

u,v

/2σ

)

u,v

−e

−σ

where u and v deﬁne the orientation and scale of the

Gabor kernels respectively, and the wave vector is de-

ﬁned as:

u,v

= k

iφ

where k

= k

max

/ f

, φ

= uπ/8; k

max

is the maximum

frequency, σ is the relative width of the Gaussian en-

velop, and f is the spacing factor between kernels

in the frequency domain (Liu and Wechsler, 2002).

The discrete ﬁlter bank of 5 different spatial frequen-

cies (v ∈ [0,··· , 4]) and 8 orientations (u ∈ [0,··· , 7])

is mostly exploited to ﬁlter face images to facilitate

multi-scale analysis for face recognition.

From the Gabor face representation, we then need

to form face features for identiﬁcation. The most pop-

ular way is to extract the LBP type patterns from the

complex Gabor transformed image. As in (Zhang

et al., 2005), the LGBP feature is extracted from the

amplitude spectrum. In (Zhang et al., 2007) and (Xie

et al., 2010), 4-ray phase-quadrant demodulator is ap-

plied to demodulate the phase from each of the com-

plex Gabor coefﬁcients, and local binary phase de-

scriptors are subsequently generated from the demod-

ulated phase spectrum. Dimension reduction can also

be used for feature construction. As in (Xie et al.,

2010), FLD is applied to form local Gabor features.

Besides, face identiﬁcation can be built on other types

of local representations, exempliﬁed as GOM (Chai

et al., 2014) and SLF (Yang et al., 2013).

Fusing other features that are independent of the

local Gabor features can also lead to better perfor-

mance: (Tan and Triggs, 2007; Su et al., 2009; Zhang

et al., 2007) fuses the global (holistic) features with

local ones at feature level; (Xie et al., 2010) proposes

fusion of Gabor phase and amplitude on the score and

feature levels; (Chai et al., 2014) fuses real, imagi-

nary, amplitude and phase. Alternatively, attaching

an illumination normalization step and weighting the

local Gabor features is shown to be helpful as well

(Cament et al., 2014).

Another face representation trend is to utilize

comparatively and/or exclusively trained CNNs to

learn discriminative metrics and features.With re-

cently launched hardware platforms (Jia et al., 2014;

Vedaldi and Lenc, 2014) and especially public acces-

sible large-scale dataset (Yi et al., 2014), developing

deep learning based face recognition approaches be-

comes feasible with less resources. Many deep net-

work architectures have been proposed as in (Schroff

et al., 2015; Parkhi et al., 2015; Yi et al., 2014).

3 THE GABOR PHASE BLOCK

MATCHING APPROACH

In this section, we ﬁrst introduce the philosophy of

our proposed approach in Subsection 3.1, and then

describe the signal representation selected for our ap-

proach and the details of the approach in Subsection

3.2.

3.1 Overview

Repeatable features extracted from small face por-

tions are known as good discriminative traits for iden-

tifying persons. In addition, such local features are

less likely to be inﬂuenced than the holistic features

by pose changes and facial expressions. Thus, it is

natural to divide face image into blocks and performs

similarity measurements between them.

Even if being different in how to construct local

features from either Gabor amplitude or phase rep-

resentation, state-of-the-art face recognition methods

do share a common implication: the spatially corre-

sponding patches/features are the best match (since

matching is only performed between spatially corre-

sponding features). This implication is hardly true

because of the movement of facial components, head

pose variablity and imprecise alignment, the spatially

corresponding patches easily become dislocated (see

Fig. 2 in (Zou et al., 2007)). In addition, matching

the spatially corresponding Gabor phase patches is

even worse than comparing Gabor amplitude patches

since phase is very sensitive to spatial shift of fa-

cial textures. To handle this problem, some solutions

deploy elastic matching strategy to allow each seg-

mented patch of one image to search the best match-

ing from spatially neighboring locations on the other

image. This has achieved better robustness to spatial

Leveraging Gabor Phase for Face Identiﬁcation in Controlled Scenarios

shift of the textures as in (Lades et al., 1993; Wiskott

et al., 1997; Hua and Akbarzadeh, 2009; Zhong and

Li, 2014).

It is not surprising that different facial area

(blocks) have different utilities for identifying peo-

ple. For instance, patches containing eyes would

have higher discrimination power than other patches.

But, it is also straightforward that the discriminative

patches could locate at any position on the face when

matching faces. This is because the discriminative

features are totally matching-pair speciﬁc. For exam-

ple, patches from the cheek area with almost no tex-

ture would be useful only when matched to a cheek

patch has scar or a mole; it won’t contribute much

when matched to similar texture-less ones. Thus, to

improve discrimination, it is necessary to weight the

corresponding features based on their utility.

In this work, we show how to use the Blocking

Matching method (Zhong and Li, 2014) to construct

local face features and perform matching between

best-matching features. Our Block Matching scheme

explicitly handle spatial shift between two patches, so

that high-deﬁnition

Gabor phase information can be

maximally utilized and patch-wise utility is evaluated

on-line in the same matching process as well. We

demonstrate the combination of Gabor phase with the

Blocking Matching is powerful in handling the hard

factors, like light, pose, facial expressions, and aging

involved in face identiﬁcation tasks.

3.2 Algorithm

The matching process of our Gabor Phase Block

Matching (GPBM) approach is illustrated in Fig.1 and

the details of the Block Matching searching scheme is

shown in Fig.2.

To match a gallery and a probe image, our GPBM

approach consists of two steps. In the ﬁrst step, both

probe and gallery face images are ﬁltered using a Ga-

bor ﬁlter bank (in the following we show that a single-

scale Gabor ﬁlter with two orientations is sufﬁcient in

our approach). The ﬁltered images are demodulated

by a Gray-coded Phase Shift Keying (PSK) demod-

ulator for smooth phase quantization. In the second

step, we divide phase images into non-overlapping

blocks. It is a natural choice to employ such a block

based approach since it is our way to construct fea-

tures. As mentioned in (Xie et al., 2010), the block

(patch) based strategy is an effective tool to handle

the so called “curse of dimensionality” and has been

proved to be an effective tool in face recognition.

Compared to approaches used quadratic phases, e.g., in

(Xie et al., 2010).

In our approach, the Block Matching method is

used to form features for identiﬁcation. The demod-

ulated phase spectra are input to the Block Match-

ing method (Zhong and Li, 2014) to form features

and calculate the the pair wise distance between a

probe (pb) image and a gallery (gl) image. Speciﬁ-

cally (Step 2 of Fig. 1), we ﬁrst segment the probe

phase spectrum into N non-overlapping patches and

the patches {f

(pb)

}

N−1

are simply formed by the raw

phase codes of the patches. For each probe patch

(pb)

centered at image coordinate (x

) (denoted

as f

(pb)

)), it searches its best matching block

within the corresponding search window and yields a

patch-wise distance vector d

denoted as:

= {d

},i ∈ [0, L −1] (1)

where L is the number of candidate gallery patches

within the (2R+1)×(2C +1) search window, i.e. L =

(2R +1) ·(2C +1) when applying full search method,

R and C stands for the searching offset in vertical and

horizontal directions respectively. Each element in d

is computed by performing an explicit matching over

the raw demodulated phase as:



XOR( f

(pb)

), f

(gl)

))

decimal



, (2)

where the patch-wise distance metric is the l2-norm

of element wise Humming distance in decimal and

(gl)

) denotes the patch that centered at image

coordinate (x

) within the search window on the

gallery face image so that,

(

= x

+ ∆x, ∆x ∈ [−C,C]

= y

+ ∆y, ∆y ∈ [−R,R].

(3)

We then evaluate the patch-wise utility by apply-

ing linear regression to the matching distance values.

This is achieved by calculating the slope k

of the lin-

ear ﬁtting of the ﬁrst 5 ascendingly sorted values of d

for normalization of the patch wise distance for each

patch, such that the normalization factor s

is calcu-

lated as:

= k

∗

, (4)

where d

∗

= min(d

). s

is then normalized by its l1-

norm as:

∗

= s

N−1

∑

n=0

. (5)

Finally, the distance between a matching pair of

probe and gallery face image is the weighted sum of

∗

as:

dist

(pb,gl)

N−1

∑

n=0

∗

·d

∗

. (6)

It is noteworthy that 1) feature extraction is not

carried out through the Block Matching process; 2)

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

Figure 1: Gabor Phase Matching using the Block Matching Method.

Figure 2: Block Matching Scheme.

the patch-wsie utility evaluation process is realized

only based on the two matching images at hand and

learner is not involved.

4 EXPERIMENTS AND RESULTS

4.1 Database Selection

There are a variety of large-scale datasets available

for benchmark evaluation of different face recognition

approaches, such as the FERET (Phillips et al., 2000),

FRGC2.0(Phillips et al., 2005) and the LFW dataset.

Since we focus on face recognition in controlled sce-

narios in this paper, the FERET database — the most

commonly used face identiﬁcation benchmark — is

selected to evaluate and compare our method with

state-of-the-art face identiﬁcation approaches. In ad-

dition, the CMU-PIE (Sim et al., 2002) dataset is

selected to evaluate our GPBM against variations of

pose, expression and illumination.

4.2 Experimental Setup

Face images are ﬁrst normalized (aligned) based on

the positions of both eyes as in (Xie et al., 2010).

A central facial area of 150 ×136 , which maintains

Figure 3: Phase codeword generator: 16-PSK demodulator

constellation.

the same height/width ratio (1.1 : 1) as in (Xie et al.,

2010; Zhang et al., 2010), is segmented from the face

image and used for our experiments.

Due to our Block Matching scheme, the Gabor

phase information with a higher deﬁnition can be uti-

lized in our approach. We found that a single-scale

Gabor ﬁlter pair with two orientations is sufﬁcient

for face identiﬁcation. In our implementation, the

selected Gabor ﬁlters have the following parameters:

v = 0, u ∈ {2, 6}, f =

√

2, k

max

= π/2, σ = 2π.

One can see that the chosen Gabor ﬁlters

have broad high-frequency coverage. These high-

frequency components correspond to facial texture

variations and are insensitive to the factors of lighting,

pose, and aging. Accordingly, to retain high phase

deﬁnition and to be tolerant to potential phase change

caused by texture shift, a Gray-coded 16-PSK demod-

ulator is used for phase demodulation and the constel-

lation is shown in Fig.3. Compared to the quadrature

phase demodulation used in (Zhang et al., 2007; Xie

et al., 2010), 4 times the phase information can be uti-

lized thanks to the employment of our block-matching

approach. With Gabor phase information, the block

matching approach is used to form features for face

identiﬁcation.

Leveraging Gabor Phase for Face Identiﬁcation in Controlled Scenarios

4.3 Evaluations on the CMU-PIE

Database

The CMU-PIE database contains 41368 images of 68

subjects. Images with Pose Label 05, 07, 09, 27, and

29 under 21 illuminations (Flash 2 to 22) of all the 68

persons are selected as the probe set.

When applying the blocking matching method,

the most important parameters are the block size (H

and W) and searching offset (R and C). We have con-

ducted a set of empirical tests over other datasets to

select suitable parameters. We found that it makes

sense to divide a central facial area into 5 ×7 batches,

which correspond to semantic facial macro features,

like eyes, nose, etc. Thus, for a facial area of 150 ×

136, a reasonable size of a block is 30 ×20. In our

implementation, we select the block size of 29 ×19

for convincing block searching (where we prefer the

block size with odd numbers). To have good cover-

age while keeping low computational complexity, the

searching offset is chosen as around

of the block

size and we select search offset of R = 7, C = 6 pixels

in our experiments. To test how sensitive the perfor-

mance is to the selected parameters, we selected the

ﬁrst 2000 probe images on the CMU-PIE to evaluate

the performance with the chosen parameters and other

parameters randomly selected around them. The eval-

uation results are shown in Fig. 4. From the test re-

sults one can see that the performance is rather insen-

sitive to the selection of parameters. Thus, in our ex-

periments with both CMU-PIE and FERET databases

we used the chosen parameters.

Figure 4: Recognition rates under different parameters on

PIE.

We then conduct experiments on the CMU-PIE

probe set and compare our GPBM with G LBP and

G LDP (Zhang et al., 2010). The G LBP is the Ga-

bor version LBP and the G LDP is a type of im-

proved Gabor amplitude Local Binary Pattern. The

G LDP achieved equivalent performance as LGXP

(Gabor Phase pattern) on the FERET evaluations so it

is a good reference for comparison. The comparative

rank-1 recognition rates are listed in Table 1. It can

be seen that our method is at least 3% better than the

Table 1: Comparative rank-1 recognition rates of GPBM on

the CMU-PIE database.

Method Accuracy

G LBP

∗

71%

G LDP 2nd-order

∗

72%

G LDP 3rd-order

∗

79%

G LDP 4th-order

∗

74%

GPBM 82%

∗

The recognition rates are estimated from Fig. 12a in

(Zhang et al., 2010).

G LDP, even though LDP extract much more compli-

cated patterns than the LBP from the Gabor ampli-

tude space. Utilizing the Gabor phase in the Block

Matching scheme is more effective in dealing with

pose and illumination changes than LBP-type patterns

extracted from the Gabor amplitude space.

4.4 Evaluations on the FERET

Database

The FERET database contains 1196 frontal face im-

ages in the gallery set, 1195 images with different ex-

pressions in the probe set “Fa”, 194 images with illu-

mination variations in the probe set “Fc”, 722 images

taken in later time in the “Dup1” set, and 234 images

taken at least 1 year later than the gallery set form the

hardest “Dup2” set. We faithfully follow the evalua-

tion protocol of the FERET dataset. The results of our

GPBM with other approaches using Gabor-phase are

listed in Table 2.

From Table 2 one can see that in a fair compar-

ison, when only Gabor phase is utilized for match-

ing, our GPBM is almost 12% better than LGXP on

the hardest “Dup2”; even in unfavorable comparisons,

where pre-processing, training, and fusion methods

are exploited by LMGEW//LN+LGXP and S[LGBP

Mag+LGXP], our GPBM still excles. To our best

knowledge, the method S[LGBP Mag+LGXP] —

aided by the Gabor amplitude and training procedures

— is state-of-the-art Gabor phase based method in

terms of performance on the hardest FERET “Dup2”,

and our GPBM is entirely comparable.

We also further compare our GPBM with other

state-of-the-art approaches based on other techniques

on the FERET in Table 3. From the table one can see

that all these approaches are based on Gabor features,

which indicates the Gabor ﬁlter is a very effective tool

for signal representation. Our GPBM method outper-

forms all the other approaches on the hardest “Dup2”

set and it features three advantages: 1) it enables high

deﬁnition Gabor phase to be utilized for face identiﬁ-

cation; 2) a single-scale Gabor ﬁlter with two orienta-

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

Table 2: Comparative rank-1 recognition rates of Gabor-phase based approaches on the FERET database.

Method Fb Fc Dup1 Dup2

LGBP Pha (Zhang et al., 2009) 93% 92% 65% 59%

LGBP Pha

weighted

(Zhang et al., 2009) 96% 94% 72% 69%

HGPP

weighted

(Zhang et al., 2007) 97.5% 99.5% 79.5% 77.8%

LGXP (Xie et al., 2010) 98% 100% 82% 83%

LGXP+BFLD (Xie et al., 2010) 99% 100% 92% 91%

S[LGBP Mag+LGXP] (Xie et al., 2010) 99% 100% 94% 93%

LMGEW//LN+LGXP (Cament et al., 2014) 99.9% 100% 94.7% 91.9%

GPBM 99.4% 100% 95.3% 94.9%

Table 3: Comparative Summary of Recent State-of-the-art Face Identiﬁcation Approaches.

Methods Image Size Photometric Gabor Filter Gabor Feature Space Training Data Rank-1 Accuracy

Processing Scale × Orient. Independent on FERET Dup2

LGXP (Xie et al., 2010) 88 ×80 No 5 ×8 Phase No 83%

LGBP+LGXP (Xie et al., 2010) 88×80 No 5 ×8 Amplitude + Phase No 93%

GOM (Chai et al., 2014) 160 ×128 No 5 ×8 Amplitude + Phase No 93.1%

LN+LGXP (Cament et al., 2014) 251×203 Yes 5 ×8 Phase No 91.9%

LN+LGBP (Cament et al., 2014) 251 ×203 Yes 5 ×8 Amplitude No 93.6%

SLF-RKR l2 (Yang et al., 2013) 150 ×130 No 5 ×8 Amplitude No 94.4%

GPBM, ours 150 ×136 No 1 ×2 Phase(explicit matching) Yes 94.9%

tions is sufﬁcient to generate an effective face image

representation, with 1/20 of the computational com-

plexity of other methods that utilized 40 Gabor ﬁlters;

3) further to this, it is not a learning-based face iden-

tiﬁcation method and, therefore, promises good gen-

eralization.

The computational complexity is always a big

concern. From Table 4 in (Mu et al., 2011), un-

der the image size of 128 ×128 with a 5 ×8 Gabor

ﬁlter bank, the histogram extraction of LGBP takes

around 0.45 seconds, S[LGBP Mag+LGXP] takes

0.99 seconds. Extracting GOM feature takes 0.7 sec-

onds (Chai et al., 2014). However, the “feature ex-

traction” time in our method is 0 seconds since only

the raw phase is used for matching; the demodula-

tion is the only on-line computation of the probe face,

thus, it is extremely fast. Our Matlab implementation

executes the matching of a face pair in 0.05 second

in average (Gabor ﬁltering included) on a 3.4GHz In-

tel CPU. We can therefore safely conclude that our

GPBM outperforms the best Gabor-phase based ap-

proach (S[LGBP Mag+LGXP]) in efﬁciency with a

big margin and we can also infer that the other meth-

ods in Table 3 could hardly be more efﬁcient than our

GPBM due to higher image resolution, Gabor face

dimensions, and additional photometric processing.

Here we should mention that our GPBM needs to run

block matching. Right now, we used an “exhaustive

search” strategy. Since we have just a few blocks per

probe image, matching is still fast. In future work, we

could also incorporate fast-search strategies from the

video compression ﬁeld to speed up face matching.

5 DISCUSSIONS AND

CONCLUSIONS

Before we conclude this paper, it would be inter-

esting to investigate “How good Deep Learning can

be in face recognition in controlled scenarios?”. To

answer this, we trained several CNNs with well-

known architectures of AlexNet (Krizhevsky et al.,

2012), VGG-net (Simonyan and Zisserman, 2014)

and Google’s InceptionNet (Szegedy et al., 2014) and

FaceNet (Schroff et al., 2015)), and evaluated them

on the most difﬁcult probe set “Dup2” of FERET

database. For fair comparisons on different architec-

tures, layers after the last spatial pooling in our imple-

mentation of the InceptionNet and the FaceNet were

replaced by two concatenated Fully Connected (FC)

layers . We used WebFace dataset (Yi et al., 2014) to

train our networks and carefully ﬁne tuned the trained

nets afterwords with FERET gallery images.

To illustrate how architecture choice affects recog-

nition performance, we investigated how the Rank-1

accuracy varies under different sizes of the FC lay-

ers. The results are enlisted in Table 4. We can see

that the architecture (length of FC) does inﬂuence the

recognition accuracy. On the one hand, explicitly in-

herit networks designed for other image classiﬁcation

Leveraging Gabor Phase for Face Identiﬁcation in Controlled Scenarios

Table 4: Rank-1 accuracy of several well-known CNN ar-

chitectures on FERET Dup2 (Input image size to CNNs is

120 ×120).

Archetectures

Length of the Last 2 FCs

FC-4096 FC-1024

AlexNet 91.9% 94.4%

VGG-13 layers 93.6% 94.9%

VGG-16 layers 93.2% 97.0%

InceptionNet 95.3% 98.7%

FaceNet 94.9% 98.3%

tasks may not perform well in novel face recognition

tasks (compare the left column to the right); investiga-

tions on suitable deep feature representations must be

made correspondingly ( here we found that FC −1024

is a good choice which is also veriﬁed by (Parkhi

et al., 2015)). On the other hand, the performance

strongly correlates to architecture in general: even

with FC −1024 the InceptionNet outperformed oth-

ers. While it is not astonishing that some CNNs out-

performed the proposed approach for almost 4%, we

can see that such advantage is not statistically signiﬁ-

cant: the best CNN correctly identiﬁed 9 more probe

faces than our proposal which made 222 correct an-

swers out of 234 probes on the “Dup2” set. One can

expect even higher accuracy from better CNN solu-

tions, but under limited conditions where CNN based

solutions are not feasible (e.g., due to lack of train-

ing data) the proposed method is still a good alter-

native due to its comparable effectiveness and high

efﬁciency.

The Gabor based solutions share a common lim-

itation. They extract low level features from image

texture through a shallow ﬁlter bank (typically of 40

ﬁlter kernels). When it comes to matching faces with

dramatic texture changes, the features are not “deep”

enough to construct discriminative representations.

This makes Gabor based solutions unsuitable for face

recognition in uncontrolled scenarios where the CNN

solutions, which build high-level features from low

level Gabor-like ﬁlters, demonstrated overwhelming

advantages.

To conclude, we propose a plain approach to lever-

age the demodulated Gabor phase for face identiﬁca-

tion based on the Block Matching method. The pro-

posed approach neither utilizes a large Gabor ﬁlter

bank nor a training process. It only depends on the

signal representation from a single-scale Gabor ﬁlter

pair to perform explicit matching over the raw Gabor

phase spectrum.

Comparative experiments show that: 1) our ap-

proach features the highest accuracy utilizing the Ga-

bor phase for face recognition; 2) our approach retains

very low computational complexity yet with compa-

rable performance to other state of the art methods

including Deep Learning methods when it works on

face recognition in controlled scenarios. Our experi-

ments demonstrate that our Block Matching method

is a powerful tool that can leverage the power of Ga-

bor phase to boost the face recognition performance.

REFERENCES

Ahonen, T., Hadid, A., and Pietikinen, M. (2004). Face

recognition with local binary patterns. In Pajdla, T.

and Matas, J., editors, Computer Vision - ECCV 2004,

volume 3021 of Lecture Notes in Computer Science,

pages 469–481. Springer Berlin Heidelberg.

Belhumeur, P. N., Hespanha, J. P., and Kriegman, D. (1997).

Eigenfaces vs. ﬁsherfaces: Recognition using class

speciﬁc linear projection. Pattern Analysis and Ma-

chine Intelligence, IEEE Transactions on, 19(7):711–

720.

Cament, L. A., Castillo, L. E., Perez, J. P., Galdames, F. J.,

and Perez, C. A. (2014). Fusion of local normalization

and gabor entropy weighted features for face identiﬁ-

cation. Pattern Recognition, 47(2):568–577.

Cao, X., Wei, Y., Wen, F., and Sun, J. (2014). Face align-

ment by explicit shape regression. International Jour-

nal of Computer Vision, 107(2):177–190.

Chai, Z., Sun, Z., Mendez-Vazquez, H., He, R., and Tan,

T. (2014). Gabor ordinal measures for face recogni-

tion. Information Forensics and Security, IEEE Trans-

actions on, 9(1):14–26.

Chan, C. H., Tahir, M. A., Kittler, J., and Pietikainen,

M. (2013). Multiscale local phase quantization for

robust component-based face recognition using ker-

nel fusion of multiple descriptors. Pattern Analy-

sis and Machine Intelligence, IEEE Transactions on,

35(5):1164–1177.

Chen, D., Cao, X., Wang, L., Wen, F., and Sun, J. (2012).

Bayesian face revisited: A joint formulation. In Com-

puter Vision–ECCV 2012, pages 566–579. Springer.

Dalal, N. and Triggs, B. (2005). Histograms of oriented gra-

dients for human detection. In Computer Vision and

Pattern Recognition, 2005. CVPR 2005. IEEE Com-

puter Society Conference on, volume 1, pages 886–

893. IEEE.

Daugman, J. (2004). How iris recognition works. Circuits

and Systems for Video Technology, IEEE Transactions

on, 14(1):21–30.

Daugman, J. (2006). Probing the uniqueness and random-

ness of iriscodes: Results from 200 billion iris pair

comparisons. Proceedings of the IEEE, 94(11):1927–

1935.

Gao, Y., Wang, Y., Zhu, X., Feng, X., and Zhou, X. (2006).

Weighted gabor features in unitary space for face

recognition. In Automatic Face and Gesture Recog-

nition, 2006. FGR 2006. 7th International Conference

on, pages 6–pp. IEEE.

VISAPP 2016 - International Conference on Computer Vision Theory and Applications

Givens, G. H., Beveridge, J. R., Lui, Y. M., Bolme, D. S.,

Draper, B. A., and Phillips, P. J. (2013). Biomet-

ric face recognition: from classical statistics to future

challenges. Wiley Interdisciplinary Reviews: Compu-

tational Statistics, 5(4):288–308.

Hua, G. and Akbarzadeh, A. (2009). A robust elastic and

partial matching metric for face recognition. In Com-

puter Vision, 2009 IEEE 12th International Confer-

ence on, pages 2082–2089. IEEE.

Huang, G., Lee, H., and Learned-Miller, E. (2012). Learn-

ing hierarchical representations for face veriﬁcation

with convolutional deep belief networks. In Computer

Vision and Pattern Recognition (CVPR), 2012 IEEE

Conference on, pages 2518–2525.

Huang, G. B., Ramesh, M., Berg, T., and Learned-Miller,

E. (2007). Labeled faces in the wild: A database for

studying face recognition in unconstrained environ-

ments. Technical Report 07-49, University of Mas-

sachusetts, Amherst.

Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J.,

Girshick, R., Guadarrama, S., and Darrell, T. (2014).

Caffe: Convolutional architecture for fast feature em-

bedding. arXiv preprint arXiv:1408.5093.

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

agenet classiﬁcation with deep convolutional neural

networks. In Pereira, F., Burges, C., Bottou, L., and

Weinberger, K., editors, Advances in Neural Informa-

tion Processing Systems 25, pages 1097–1105. Curran

Associates, Inc.

Lades, M., Vorbruggen, J. C., Buhmann, J., Lange, J.,

von der Malsburg, C., Wurtz, R. P., and Konen, W.

(1993). Distortion invariant object recognition in the

dynamic link architecture. Computers, IEEE Transac-

tions on, 42(3):300–311.

Li, H., Hua, G., Lin, Z., Brandt, J., and Yang, J. (2013).

Probabilistic elastic matching for pose variant face

veriﬁcation. In Computer Vision and Pattern Recogni-

tion (CVPR), 2013 IEEE Conference on, pages 3499–

3506. IEEE.

Liu, C. and Wechsler, H. (2002). Gabor feature based classi-

ﬁcation using the enhanced ﬁsher linear discriminant

model for face recognition. Image processing, IEEE

Transactions on, 11(4):467–476.

Lowe, D. G. (2004). Distinctive image features from scale-

invariant keypoints. International journal of computer

vision, 60(2):91–110.

Mu, M., Ruan, Q., and Guo, S. (2011). Shift and gray

scale invariant features for palmprint identiﬁcation us-

ing complex directional wavelet and local binary pat-

tern. Neurocomputing, 74(17):3351–3360.

Oppenheim, A. V. and Lim, J. S. (1981). The importance of

phase in signals. Proceedings of the IEEE, 69(5):529–

541.

Parkhi, O. M., Vedaldi, A., and Zisserman, A. (2015). Deep

face recognition. Proceedings of the British Machine

Vision Conference.

Phillips, P. J., Flynn, P. J., Scruggs, T., Bowyer, K. W.,

Chang, J., Hoffman, K., Marques, J., Min, J., and

Worek, W. (2005). Overview of the face recogni-

tion grand challenge. In Computer vision and pattern

recognition, 2005. CVPR 2005. IEEE computer soci-

ety conference on, volume 1, pages 947–954. IEEE.

Phillips, P. J., Moon, H., Rizvi, S. A., and Rauss, P. J.

(2000). The feret evaluation methodology for face-

recognition algorithms. Pattern Analysis and Machine

Intelligence, IEEE Transactions on, 22(10):1090–

1104.

Pinto, N., DiCarlo, J. J., and Cox, D. D. (2009). How far can

you get with a modern face recognition test set using

only simple features? In Computer Vision and Pattern

Recognition, 2009. CVPR 2009. IEEE Conference on,

pages 2591–2598. IEEE.

Schroff, F., Kalenichenko, D., and Philbin, J. (2015).

Facenet: A uniﬁed embedding for face recognition

and clustering. In Proceedings of the IEEE Confer-

ence on Computer Vision and Pattern Recognition,

pages 815–823.

Sim, T., Baker, S., and Bsat, M. (2002). The cmu pose,

illumination, and expression (pie) database. In Auto-

matic Face and Gesture Recognition, 2002. Proceed-

ings. Fifth IEEE International Conference on, pages

46–51. IEEE.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

CoRR, abs/1409.1556.

Su, Y., Shan, S., Chen, X., and Gao, W. (2009). Hierarchi-

cal ensemble of global and local classiﬁers for face

recognition. Image Processing, IEEE Transactions

on, 18(8):1885–1896.

Sun, Y., Wang, X., and Tang, X. (2013). Deep learning face

representation from predicting 10,000 classes. In Pro-

ceedings of the IEEE Conference on Computer Vision

and Pattern Recognition, pages 1891–1898.

Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,

Anguelov, D., Erhan, D., Vanhoucke, V., and Rabi-

novich, A. (2014). Going deeper with convolutions.

Computer Vision, 2014 IEEE 12th International Con-

ference on.

Taigman, Y., Yang, M., Ranzato, M., and Wolf, L. (2013).

Deepface: Closing the gap to human-level perfor-

mance in face veriﬁcation. In Proceedings of the IEEE

Conference on Computer Vision and Pattern Recogni-

tion, pages 1701–1708.

Tan, X. and Triggs, B. (2007). Fusing gabor and lbp fea-

ture sets for kernel-based face recognition. In Analysis

and Modeling of Faces and Gestures, pages 235–249.

Springer.

Turk, M. A. and Pentland, A. P. (1991). Face recogni-

tion using eigenfaces. In Computer Vision and Pat-

tern Recognition, 1991. Proceedings CVPR’91., IEEE

Computer Society Conference on, pages 586–591.

IEEE.

Vedaldi, A. and Lenc, K. (2014). Matconvnet-

convolutional neural networks for matlab. arXiv

preprint arXiv:1412.4564.

Wiskott, L., Fellous, J.-M., Kuiger, N., and Von Der Mals-

burg, C. (1997). Face recognition by elastic bunch

graph matching. Pattern Analysis and Machine Intel-

ligence, IEEE Transactions on, 19(7):775–779.

Leveraging Gabor Phase for Face Identiﬁcation in Controlled Scenarios

Wright, J., Yang, A. Y., Ganesh, A., Sastry, S. S., and Ma,

Y. (2009). Robust face recognition via sparse repre-

sentation. Pattern Analysis and Machine Intelligence,

IEEE Transactions on, 31(2):210–227.

Xie, S., Shan, S., Chen, X., and Chen, J. (2010). Fus-

ing local patterns of gabor magnitude and phase for

face recognition. Image Processing, IEEE Transac-

tions on, 19(5):1349–1361.

Yang, M., Zhang, L., Shiu, S.-K., and Zhang, D. (2013).

Robust kernel representation with statistical local fea-

tures for face recognition. Neural Networks and

Learning Systems, IEEE Transactions on, 24(6):900–

912.

Yi, D., Lei, Z., and Li, S. Z. (2013). Towards pose ro-

bust face recognition. In Computer Vision and Pat-

tern Recognition (CVPR), 2013 IEEE Conference on,

pages 3539–3545. IEEE.

Yi, D., Lei, Z., Liao, S., and Li, S. Z. (2014). Learn-

ing face representation from scratch. arXiv preprint

arXiv:1411.7923.

Zhang, B., Gao, Y., Zhao, S., and Liu, J. (2010). Lo-

cal derivative pattern versus local binary pattern: face

recognition with high-order local pattern descriptor.

Image Processing, IEEE Transactions on, 19(2):533–

544.

Zhang, B., Shan, S., Chen, X., and Gao, W. (2007). His-

togram of gabor phase patterns (hgpp): A novel object

representation approach for face recognition. Image

Processing, IEEE Transactions on, 16(1):57–68.

Zhang, W., Shan, S., Gao, W., Chen, X., and Zhang, H.

(2005). Local gabor binary pattern histogram se-

quence (lgbphs): A novel non-statistical model for

face representation and recognition. In Computer

Vision, 2005. ICCV 2005. Tenth IEEE International

Conference on, volume 1, pages 786–791. IEEE.

Zhang, W., Shan, S., Qing, L., Chen, X., and Gao, W.

(2009). Are gabor phases really useless for face recog-

nition? Pattern Analysis and Applications, 12(3):301–

307.

Zhong, Y. and Li, H. (2014). Is block matching an alterna-

tive tool to lbp for face recognition? In Image Pro-

cessing (ICIP), 2014 IEEE International Conference

on, pages 723–727.

Zou, J., Ji, Q., and Nagy, G. (2007). A comparative study

of local matching approach for face recognition. Im-

age Processing, IEEE Transactions on, 16(10):2617–

2628.

VISAPP 2016 - International Conference on Computer Vision Theory and Applications