Paper Recent Advances in Biometric Recognition

Paper

Recent Advances in Biometric Recognition

Koichi Ito (member)†, Takafumi Aoki (member)†

Abstract This paper presents recent advances in biometric recognition, where we focus on face, fingerprint and iris

recognition, which are major research topics on biometric recognition. We summarize the research trend of face, fingerprint

and iris recognition over the past decade. This paper also presents our activities of biometric recognition. Our approach

employs the phase information obtained by Discrete Fourier Transform (DFT) of images. The phase information preserves

the inherent features of the image, and its correlation function, called phase correlation or Phase-Only Correlation (POC),

gives us both the good similarity measure for biometric recognition and the translational displacement for image registration.

Our approach of using phase information has been successfully applied to fingerprint, face, iris, palmprint, finger knuckle and

dental recognition. Among them, we present some interesting results of palmprint recognition, finger knuckle recognition and

dental recognition.

Key words: biometrics, face recognition, fingerprint recognition, iris recognition, palmprint recognition, finger knuckle recognition,

disaster victim identification, phase-only correlation

1. Introduction

Biometric authentication (or simply biometrics) is to

identify a person based on the physiological or behav-

ioral characteristics1)2) such as fingerprint, face, iris,

voice, signature, etc. Biometrics has attracted exten-

sive attention as a new authentication approach against

traditional ones such as key, password, etc. Biometric

traits are not stolen and forgotten compared with key,

card and password. Therefore, biometrics techniques

provide us better security and greater convenience than

traditional person authentication techniques. Practical

person authentication systems using fingerprint, face,

iris, etc. have been commercially available and used in

access control, ATM, etc.

Jain et al.3) summarized what biological measure-

ments qualify to be a biometric trait. They introduced

the following requirements to use physiological or be-

havioral characteristic as a biometric trait:

• Universality: each person should have the charac-

teristic.

• Distinctiveness: any two persons should be suffi-

ciently different in terms of the characteristic.

• Permanence: the characteristic should be suffi-

ciently invariant over a period of time.

• Collectability: the characteristic can be measured

Received ; Revised ; Accepted

†Graduate School of Information Sciences, Tohoku University

(6-6-05, Aramaki Aza Aoba, Sendai, 980-8579, Japan.)

quantitatively.

They also presented some issues to be considered in a

practical biometric system:

• Performance, which refers to the achievable recog-

nition accuracy and speed, the resources required to

achieve the desired recognition accuracy and speed, as

well as the operational and environmental factors that

affect the accuracy and speed.

• Acceptability, which indicates the extent to which

people are willing to accept the use of a particular bio-

metric trait in their daily lives,

• Circumvention, which reflects how easily the sys-

tem can be fooled using fraudulent methods.

Biometric techniques to be used in the practical sys-

tem depend heavily on application requirements. The

above seven factors can be used to compare biomet-

ric techniques as shown in Table 1, which was derived

on the perception of Jain et al3). This table provides

us good suggestions when we consider which biomet-

ric techniques are selected in practice. Let us focus on

performance and collectability of biometric traits. A

face has low performance due to weakness against envi-

ronmental variations, while it is easy to capture a face

by a camera. An iris has high performance due to its

distinctive texture pattern, while an iris image is cap-

tured by a special imaging device. A face is suitable

for low-security level application such as PC login be-

cause of its high collectability. On the other hand, an

iris is suitable for high-security level application such as

immigration control because of its high performance, al-

ITE Trans. on MTA Vol. 6, No. 1, pp. 64-80 (2018)

64

Copyright © 2018 by ITE Transactions on Media Technology and Applications (MTA)

Received July 19, 2017; Revised October 4, 2017; Accepted October 11,2017

Table 1 Comparison of various biometric techniques on the perception of Jain et al.3), where H, M and

L indicate High, Middle and Low, respectively.

Universality Distinctiveness Permanence Collectability Performance Acceptability Circumvention

DNA H H H L H L L

Ear M M H M M H M

Face H L M H L H H

Facial thermogram H H L H M H L

Fingerprint M H H M H M M

Gait M L L H L H M

Hand geometry M M M H M M M

Hand vein M M M M M M L

Iris H H H M H L L

Keystroke L L L M L M M

Odor H H H L L M L

Palmprint M H H M H M M

Retina H H M L H L L

Signature L L L H L H H

Voice M L L M L H H

though a special imaging device is required to capture

iris images. Performance can be complemented with

combining multiple biometric traits such as face and

iris to keep both high performance and high collectabil-

ity. This approach is known as multimodal biometrics4),

which is one of active research topics in biometrics.

Table 2 summarizes the number of papers for each

biometric trait presented in the international confer-

ences related to biometrics, where the number of pa-

pers was counted by the authors. This summary pro-

vides us the research trends in the field of biometrics

over the past 13 years. Researches on face, fingerprint

and iris are always hot, since the number of their pa-

pers is constantly large and is always more than other

biometric traits. A face is a major research topic in

many fields such as computer vision, pattern recog-

nition, image processing and biometrics. A variety

of face image processing methods has been proposed,

since the performance of face image processing is sig-

nificantly influenced by environmental changes such as

head pose, expression and illumination changes. A new

research topic on fingerprint recognition has been ex-

plored, since minutiae-based matching exhibits suffi-

cient performance on fingerprint recognition and practi-

cal fingerprint recognition systems have been developed.

Latent fingerprint recognition is one of the new topics,

which requires new preprocessing methods such as fin-

gerprint segmentation, ridge enhancement and minutiae

extraction specially designed for latent fingerprint im-

ages. A new research topic on iris recognition has been

explored as well as fingerprint recognition, since iriscode

is the first choice of iris recognition because of its high

recognition performance. The purpose of iris recogni-

tion is changed from a person to pedestrians from the

viewpoint of surveillance applications. Therefore, oc-

ular recognition, which uses the surrounding region of

the eye for biometric recognition, is considered as a new

topic instead of iris recognition, since iris recognition at

a distance is a difficult problem.

Figure 1 shows a flow diagram of a standard bio-

metrics system. Note that we assume an image-based

system for a brief description in the following. This sys-

tem consists of 5 components: (i) sensing, (ii) prepro-

cessing, (iii) feature extraction, (iv) database and (v)

matching. In the sensing step, an image of a biometric

trait is captured using a sensor. For example, a cam-

era is used in face, palmprint, finger knuckle and gait

recognition and a special sensor is used in iris, signature

and vein recognition. Preprocessing consists of a set

of image processing methods such as contrast enhance-

ment, noise removal, geometric transformation, Region

Of Interest (ROI) extraction, etc. The performance of

preprocessing is important for the subsequent step of

feature extraction, since the captured image usually in-

cludes unnecessary background components for biomet-

ric recognition. In the feature extraction step, features

to be matched are extracted from an ROI image, which

is the most active topic in biometric recognition. Local

features are designed depending on the type of biomet-

ric traits in the most cases. Database stores registered

features for the matching step. In the matching step,

a similarity or a dissimilarity between registered and

input features is calculated to make a final decision.

Each component is a major research topic on biomet-

rics. In addition, there are other research topics con-

sidered in the biometric system such as anti-spoofing,

template protection, cancelable biometrics and multi-

modal biometrics from the viewpoint of system security.

As mentioned above, biometrics is a kind of multidisci-

plinary research fields and a variety of methods of com-

65

Invited Paper » Recent Advances in Biometric Recognition

Table 2 Research trends in the international conferences on biometrics, where the number of papers

provides us the research trends in the field of biometrics. ICBA: International Conference on

Biometric Authentication, ICB: International Conference on Biometrics, IJCB: International

Joint Conference on Biometrics.

ICBA 2004 ICB 2006 ICB 2007 ICB 2009 IJCB 2011 ICB 2012 ICB 2013 IJCB 2014 ICB 2015 ICB 2016

Face 30 27 41 44 44 25 24 21 27 18

Voice 8 3 6 10 2 3 1 0 1 0

Fingerprint 23 19 21 11 14 11 11 14 14 13

Palm 3 2 4 6 3 3 2 1 5 2

Multimodal 10 7 8 24 16 4 1 3 1 0

Gait 0 3 5 6 7 2 2 2 2 3

Iris 11 18 12 12 6 16 13 16 7 9

Signature 13 4 10 4 4 0 5 1 3 0

Others 6 21 17 8 14 17 13 18 11 7

puter vision, pattern recognition and image processing

techniques are required to develop a reliable and high-

performance biometrics system, although biometrics is

essentially a pattern recognition problem.

This paper presents recent advances in biometric

recognition, where we focus on face, fingerprint and

iris recognition, which are major research topics on bio-

metric recognition. We summarize the research trend

of face, fingerprint and iris recognition over the past

decade. This paper also presents our activities of bio-

metric recognition. Our approach employs the phase

information obtained by Discrete Fourier Transform

(DFT) of images. The phase information preserves the

inherent features of the image, and its correlation func-

tion, called phase correlation or Phase-Only Correla-

tion (POC), gives us both the good similarity measure

for biometric recognition and the translational displace-

ment for image registration. Our approach of using

phase information has been successfully applied to fin-

gerprint, face, iris, palmprint, finger knuckle and dental

recognition. We provide a brief introduction of our re-

search results of palmprint recognition, finger knuckle

recognition and dental recognition, which are interest-

ing, practical and useful in the field of biometrics.

2. Face Recognition

This section describes the research trend in face

recognition. Figure 2 shows a standard flow diagram

of face recognition systems, which consists of 4 steps:

(i) face detection, (ii) normalization, (iii) feature ex-

traction and (iv) matching. We summarize the techni-

cal advances in each step and present recent research

topics on face recognition.

2. 1 Face Detection

Face detection, which is the first process of face recog-

nition, extracts a face region from an input image. The

accuracy of face detection is important especially for

face recognition at a distance such as surveillance ap-

plication, since there are multiple faces with different

size in an image captured by a surveillance camera.

The most famous method was proposed by Viola et

al.5), which is also called the Viola-Jones method. This

method first extracts Haar-like features from an image,

where the integral image is used for fast Haar-like fea-

ture extraction. Next, a variant of AdaBoost is used

to select the best features and to train classifiers. A

strong classifier is obtained by constructing a cascade of

weak classifiers to boost the classification performance

of simple classifiers. In the training, a huge number of

face and non-face images are required to make a good

face detector. The Viola-Jones method with trained

classifiers is available in OpenCV∗, which is a famous

computer vision library. The OpenCV implementation

of the Viola-Jones method is a de-facto standard, since

users do not need a time-consuming training to make a

face detector.

The use of the Viola-Jones method makes it possi-

ble to detect near frontal faces from an image. Face

detection in real-world applications has to take into ac-

count the unconstrained conditions such as large pose

and expression changes, large occlusions, illumination

changes, etc. and is still one of the most studied topics

in computer vision∗∗. We introduce one of the state-

of-the-art studies of face detection. Yang et al. cre-

ated a face detection benchmark dataset, which is called

the WIDER FACE dataset∗∗∗. The WIDER FACE

dataset consists of 393,703 labeled face bounding boxes

in 32,203 images. Images in the dataset have a high de-

gree of variability in scale, pose, occlusion, expression,

appearance (makeup) and illumination. Yang et al.

proposed the baseline face detection method using Con-

volutional Neural Network (CNN)7), where this method

employs the multi-scale cascade CNN to deal with large

∗ OpenCV: http://opencv.org/∗∗ Please refer to the literature6) for the detailed survey of the re-

cent face detection methods.∗∗∗ WIDER FACE: http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace/

ITE Trans. on MTA Vol. 6, No. 1 (2018)

66

Biometric trait Sensing Preprocessing Feature

extraction Database

Matching Genuine/ImpostorBiometric trait Sensing Preprocessing Feature

extraction

Anti-spoofing

Cancelable biometricsTemplate protection

Multimodal biometrics

Enrollment

Verification

Fig. 1 Flow diagram of a standard biometrics system and research topics.

Face image

Face detection Normalization Feature

extraction MatchingGenuine

or Impostor

Database

Fig. 2 Flow diagram of a general face recognition system.

scale variations of faces. They also compared the face

detection accuracy of the proposed method with four

representative methods. Using such a large-scale face

image dataset, a lot of face detection methods using

a deep learning approach has been proposed in recent

years.

2. 2 Normalization

Face images have to be normalized in terms of head

pose and expression in order to exhibit good recogni-

tion performance. This process is important to deal

with face images under the unconstrained conditions.

In general, landmarks are detected on a face and are

used to normalize head pose and expression. One of the

famous landmark detection methods is Active Appear-

ance Model (AAM) proposed by Cootes et al.8). AAM is

a parametric face model of both landmarks and texture,

which is derived by using Principal Component Analy-

sis (PCA). The demo software is available on the web∗.AAM cannot handle a large head pose change, since

head pose changes are essentially 3D transformation.

On the other hand, a 3D face model, which is called

3D Morphable Model (3DMM), has been proposed by

Blanz et al.9). 3DMM can handle a large head pose

change, while this method requires a lot of 3D face mod-

els in the training. Nowadays, it is easy to capture 3D

face data and process them because of advances in com-

∗ aam tools: http://personalpages.manchester.ac.uk/staff/timothy.

f.cootes/software/am_tools_doc/index.html

puter technology. Large Scale Facial Model (LSFM)

has been created from a large scale dataset by Booth

et al.10), where the dataset includes face images and 3D

data captured from 9,663 subjects. LSFM is a para-

metric 3D facial model of head pose, facial expression,

age and ethnicity. The source code of LSFM is avail-

able on the web∗∗. The use of such parametric 3D face

models makes it possible to deal with faces under the

unconstrained conditions.

2. 3 Feature Extraction and Matching

Feature extraction and matching are core processes

of face recognition. The traditional methods employ

PCA, Independent Component Analysis (ICA) and

Linear Discriminant Analysis (LDA)11), where such a

feature is represented as a point on the subspace in

PCA, ICA, and LDA. PCA-based method is known

as Eigenface12) and LDA-based method is known as

Fisherface13). Other methods employ subspace meth-

ods such as CLAss-Featuring Information Compression

(CLAFIC) method14), subspace method14), mutual sub-

space method and its extensions such as constrained

mutual subspace method15), and multiple constrained

mutual subspace method16), etc., where such a fea-

ture is represented as a set of bases of the subspace.

The approaches mentioned above transform the high-

dimensional image space into the low-dimensional sub-

spaces and provide good representation and good dis-

∗∗ LSFM: https://github.com/menpo/lsfm

67


crimination for face recognition by selecting effective

subspaces. The drawback of such approaches is that

position and intensity of all the face images have to be

aligned. Therefore, these approaches may exhibit good

recognition performance only for face images captured

under the desired condition.

Recently, Local Binary Patterns (LBPs) have been

proposed and applied to face recognition17). LBP is ob-

tained by thresholding neighborhoods of each pixel with

the center pixel value, and then the histogram of LBPs

is used as a texture descriptor. So far, the improved

versions of the LBP-based method have been proposed

and been applied to various biometric recognition prob-

lems. LBP has the versatility for image matching and is

applied to solve computer vision problems18), since LBP

does not need any optimization process. On the other

hand, LBP cannot handle large deformation of images

and also may not exhibit the comparable performance

with the other methods specified to each biometric trait

due to its versatility. The original implementation of

LBPs is available on the web∗.The deep learning-based approach has a significant

impact on face recognition researchers. The face im-

age dataset called Labeled Faces in the Wild (LFW)∗∗

is known as one of difficult face image datasets, since

this dataset is designed for studying the problem of un-

constrained face recognition. Taigman et al.19) employ

CNN to extract features to be matched. Their CNN

model, which is called DeepFace, is trained using a

large-scale face dataset collected from Facebook, where

the number of face images is 4.4 million captured from

4,030 persons. The recognition accuracy of DeepFace is

97.35%, while that of human is 97.53%. The deep learn-

ing approach achieved a breakthrough in face recogni-

tion, since the state-of-the-art methods exhibited the

recognition accuracy of about 90% at that time. Vari-

ous CNN models have been proposed for face recogni-

tion since DeepFace was proposed and their recognition

accuracy is comparable to a human quality.

2. 4 Face Attributes

Face recognition can be used for practical situations

because of the advent of deep learning as mentioned

above. The state-of-the-art methods of face recognition

exhibit comparable recognition performance of human

even for face images captured under the unconstrained

conditions such as large pose and expression changes,

∗ LBP software: http://www.cse.oulu.fi/wsgi/MVG/Downloads/

LBPSoftware∗∗ LFW: http://vis-www.cs.umass.edu/lfw/

large occlusions, illumination changes, etc.

Further performance improvement of face recognition

is to use face attributes. Various characteristics, i.e.,

face attributes, include in a face such as gender, hair,

skin color, eyeglass, shape, etc. The use of such infor-

mation makes it possible to classify face images accord-

ing to categories of face attributes in advance, resulting

in improving recognition performance and reducing the

computation time. This approach is known as soft bio-

metrics. Jain et al. defined that soft biometric traits as

characteristics that provide some information about the

individual, but lack the distinctiveness and permanence

to sufficiently differentiate any two individuals20).

The initial approach of predicting face attributes em-

ploys the statistical approach such as Bayesian21). It

is difficult to design feature descriptors for predicting

attributes, since there is a lot of types of attributes in-

cluded in a face, resulting in low accuracy of prediction.

Recently, the deep learning approach is applied to pre-

dict face attributes22). The use of deep learning makes

it possible to predict precise attributes from a face.

3. Fingerprint Recognition

This section describes the research trend in finger-

print recognition. Fingerprints are the most widely de-

veloped biometric traits and are used for person au-

thentication more than 100 years ago23). The finger-

print technology has already been put to practical use

in various applications from forensics to high-security

access. New issues on fingerprint recognition have been

explored until now, although the de-facto standard fin-

gerprint recognition algorithm is available.

Figure 3 shows a standard flow diagram of fingerprint

recognition systems, which consists of 4 steps: (i) seg-

mentation, (ii) enhancement, (iii) minutiae extraction

and (iv) matching. First, the area of a fingerprint is

extracted from an input image. This segmentation can

be done by simple image processing. Next, a finger-

print image is enhanced so as to extract minutiae accu-

rately. Ridges of a fingerprint can be enhanced using a

set of Gabor filters24). The Matlab code is available

on the web∗∗∗. Binarization and thinning are applied to

the enhanced fingerprint. Minutiae are extracted using

a simple coordinate model. A pixel corresponding to

minutiae is characterized by a crossing number, which

is defined by the sum of differences between pairs of

adjacent pixels in 8-neighborhood. The most common

∗∗∗ Matlab code: http://www.peterkovesi.com/matlabfns/index.

html#fingerprints


68

feature descriptor derived from minutiae information is

a triplet consisting of minutia location coordinates and

the minutia angle. In general, feature descriptor de-

fined by the geometric relationship among neighboring

minutiae is used to enhance robustness against finger-

print deformation25). The matching score is calculated

by the distance between feature descriptors of minutiae.

We present recent research topics on fingerprint

recognition in the following.

3. 1 Fingerprint Matching

A huge number of fingerprint matching algorithms

has been developed because of the existence of pub-

lic fingerprint image datasets and evaluation protocols.

One of the most famous fingerprint datasets is pro-

vided by Fingerprint Verification Competition (FVC)∗,where FVC was held in 2000, 2002, 2004 and 2006.

The book23) includes fingerprint image datasets used

in FVC2000, FVC2002 and FVC2004. FVC has been

renewed as a web-based automated evaluation system

for fingerprint recognition algorithms, which is called

FVC-onGoing∗∗. Fingerprint recognition algorithms

made rapid growing, since academic and industrial re-

searchers competed on recognition performance of fin-

gerprint recognition algorithms through FVC. Fierrez

et al.26) summarized fingerprint recognition algorithms

submitted to FVC 2004 and consider the combination

of algorithms to improve the performance of fingerprint

recognition. Recently, a new minutia matching algo-

rithms, called Minutiae Cylinder Code (MCC), was pro-

posed by Cappelli et al.27), which is used as a base-

line algorithm in FVC-onGoing. MCC describes a local

structure of each minutia. This descriptor encodes spa-

tial and directional relationships between the minutia

and its neighborhood, which is represented as a cylinder

whose base and height are related to the spatial and di-

rectional information, respectively. The SDK of MCC

is available on the web∗∗∗.3. 2 Latent Fingerprint Recognition

Latent fingerprints obtained from crime scenes have

been used in forensic identification more than a cen-

tury. The manual intervention of experts is still re-

quired for latent fingerprint verification, while the per-

formance of Automated Fingerprint Identification Sys-

∗ FVC2000: http://bias.csr.unibo.it/fvc2000/

FVC2002: http://bias.csr.unibo.it/fvc2002/

FVC2004: http://bias.csr.unibo.it/fvc2004/

FVC2006: http://bias.csr.unibo.it/fvc2006/∗∗ FVC-onGoing: https://biolab.csr.unibo.it/FVCOnGoing/UI/Form/

Home.aspx∗∗∗ MCC SDK: http://biolab.csr.unibo.it/

tems (AFISs) has been significantly improved with the

recent development of technology. The difficulty in la-

tent fingerprint recognition is mainly due to (i) poor

quality of ridge information, (ii) small finger area and

(iii) large nonlinear deformation28). Although minutiae

matching is also used in latent fingerprint matching, it

is significantly difficult to extract minutiae from latent

fingerprint images. Therefore, preprocessing methods

for latent fingerprint images such as segmentation, en-

hancement, minutiae extraction have been mainly pro-

posed29). Latent fingerprint matching has been still an

open problem in the field of fingerprint recognition be-

cause of its difficulty. There is a good survey paper for

latent fingerprint matching30). For more details, please

refer to this paper.

3. 3 Hand-based Biometrics

A hand has a lot of biometrics traits other than a

fingerprint. There are some relatively new biometric

traits in a hand such as palmprint, finger knuckle and

vein.

A palm is a large inner surface of a hand with many

features such as principle lines, ridges, minutiae, tex-

ture, etc., and is expected to be one of the distinctive

biometric traits31)32). Unlike a fingerprint, a palm image

can be captured using a camera under unconstrained

environments, resulting in realizing a user-friendly con-

tactless biometric recognition system. One of the pio-

neer researches on palmprint recognition was reported

by Zhang et al.33). They proposed a baseline palm-

print recognition method and created a palmrpint im-

age database∗4. In addition, some practical systems

using palmrpint recognition have been proposed such

as a palmprint recognition system for mobile phones34)

and a touchless palmprint recognition system35). There

is a good survey paper for palmprint recognition32)36).

For more details, please refer to these papers.

An outer surface of a finger has three knuckles: a

distal interphalangeal (DIP) joint, a proximal interpha-

langeal (PIP) joint and a metacarpophalangeal (MCP)

joint as shown in Fig. 4. Kumar et al.37) categorized

three finger joints into major and minor finger knuckles,

where a DIP joint is a first minor finger knuckle, a PIP

joint is a major finger knuckle and an MCP joint is a

second minor finger knuckle. It is easy to capture such

patterns on a finger knuckle by a camera. This advan-

tage allows us to develop a flexible and compact bio-

metric authentication system. A finger knuckle is also

∗4 PolyU Palmprint Database: http://www4.comp.polyu.edu.hk/

~biometrics/

69


Fingerprint image Segmentation Enhancement Minutiae

extraction MatchingGenuine

orImpostor

Database

Fig. 3 Flow diagram of a general fingerprint recognition system.

Distal interphalangeal (DIP) joint

Proximal interphalangeal

(PIP) joint

Metacarpophalangeal (MCP) joint

Fig. 4 A taxonomy of finger knuckle joints: Blue-

colored circles indicate distal interphalangeal

(DIP) joints, green-colored circles indicate proxi-

mal interphalangeal (PIP) joints and red-colored

circles indicate metacarpophalangeal (MCP)

joints.

expected to be distinctive as well as a fingerprint and

a palmprint, although statistical analysis using a huge

dataset has to be required to demonstrate the unique-

ness of finger knuckle patterns37). A finger knuckle is

a relatively new biometric trait in contrast with fa-

mous biometric traits such as face, fingerprint and iris,

where one of pioneer researches has been reported in

2005 by Woodard et al.38). Some finger knuckle recog-

nition methods have been reported39) 42), where most of

literature use public dataset such as the PolyU FKP

database∗5.A vein is a blood pattern under the hand skin.

The advantage of vein patterns used in biometrics is

high distinguishability and robustness against spoof-

ing. Therefore, biometric recognition systems using

vein patterns have been used in high-security applica-

tions such as ATM and access control. There are two

∗5 PolyU FKP database: http://www4.comp.polyu.edu.hk/~biometrics/

FKP.htm

types of vein patterns such as hand vein43) and finger

vein44). Recently, the international competition of fin-

ger vein recognition∗ has been held in conjunction with

International Conference on Biometrics.

4. Iris Recognition

This section describes the research trend in iris recog-

nition45). An iris is the annular part between the pupil

and the white sclera and has a complex pattern deter-

mined by the chaotic morphogenetic processes during

embryonic development. The iris pattern is unique to

each person and to each eye and is essentially stable

over a lifetime. Furthermore, an iris image is typically

captured using a contactless imaging device, which is

of great importance in practical applications. Figure 5

shows a standard flow diagram of iris recognition sys-

tems, which consists of 4 steps: (i) iris segmentation,

(ii) iris normalization, (iii) feature extraction and (iv)

matching.

Most of the commercial iris recognition systems im-

plement the iriscode algorithm proposed by Daugman46), which is a famous iris recognition algorithm. The

standard procedure of the iriscode algorithm is briefly

described in the following. First, an image including an

eye is captured by a camera. In most cases, infrared

illumination is used in image acquisition, since it is dif-

ficult to separate a dark iris of Asians from their black

pupil. Next, the iris region is extracted from the cap-

tured image, where this step is called iris segmentation.

Then, the iris region is normalized to compensate for

the elastic deformations in iris texture by mapping pixel

values from the Cartesian coordinate system to the po-

lar coordinate system. Feature vectors are extracted by

applying Gabor filters to the normalized image. The

outputs are binarized to generate a 2Kbit iriscode. Fi-

nally, the Hamming distance between two iriscodes is

used for matching.

The iris recognition system has been put into prac-

∗ ICFVR2017: http://pkurate.org/


70

Iris image Segmentation Normalization Featureextraction Matching

Genuineor

Impostor

Database

Fig. 5 Flow diagram of a general iris recognition system.

tical use in high-level security applications due to the

high discrimination capability of iris texture. The im-

migration system using iris recognition has been already

introduced to provide automated clearance through UK

immigration. There are still unsolved problems in iris

recognition under unconstrained conditions and at a

distance, although the de-facto standard iris recognition

algorithm, i.e., iriscode, is available. We summarize re-

cent research topics on iris recognition in the following.

4. 1 Iris Segmentation

Iris segmentation is to extract the valid part of the iris

from the input image. The performance of iris recogni-

tion is heavily influenced by the accuracy of iris segmen-

tation. It is, however, difficult to perform accurate iris

segmentation under unconstrained conditions45). The

iris is often partially occluded by eyelids, eyelashes, and

shadows and is occluded by specular reflections when

the user wears glasses. The pupillary and limbic bound-

aries are noncircular. Other challenges of iris segmen-

tation include defocusing, motion blur, poor contrast,

oversaturation, etc. Iris segmentation needs to find the

pupillary and limbic boundaries of the iris, localize its

upper and lower eyelids if they occlude, and detect and

exclude any superimposed occlusions of eyelashes, shad-

ows, or reflections. The traditional approach of iris seg-

mentation employs circle fitting46), while this approach

cannot be used under unconstrained conditions. To

accurately segment the iris region depending on the

iris shape, Shar et al. used geodesic active contours

(GACs)47) and He et al. used an elastic model with

spline-based edge fitting48).

4. 2 Ocular Recognition

Iris recognition at a distance is considered to real-

ize high-level security in surveillance applications due

to its high discriminant capability compared with other

biometric traits. The competition for iris recognition

at a distance was held in 2008 by National Institute of

Standards and Technology (NIST), the United States,

which is called Multiple Biometric Grand Challenge

(MBGC)∗. The video sequence of a walking person cap-

tured with near-infrared illumination was used for iris

recognition in MBGC. There is no expected algorithm

submitted due to high difficulty in iris recognition such

as heavy motion blur, low resolution, poor texture, etc.

Addressing the above problem, an ocular image, which

is the surrounding region of the eye including the iris,

is used as a new biometric trait for the purpose of per-

son authentication at a distance with high-level secu-

rity49) 51). Person authentication using eye regions is

called ocular recognition or periocular recognition. In

the case of using ocular images, iris segmentation is not

required. Therefore, it is expected that ocular recogni-

tion can be used for recognizing pedestrians. NIST held

the competition of ocular recognition, which is called

Face and Ocular Challenge Series (FOCS)∗∗, to explore

new biometric recognition algorithms using ocular im-

ages. Although there has been some literature for ocu-

lar recognition52) 56), further improvement of algorithms

is required, since the recognition performance of these

algorithms is about 90% in the FOCS dataset.

5. Biometric Recognition Using Phase-

Only Correlation

This section presents our activities of biometric recog-

nition. We consider employing the phase information

obtained by DFT of images. The phase information

preserves the inherent features of the image, and its

correlation function, called phase correlation or POC,

gives us both the good similarity measure for biomet-

ric recognition and the translational displacement for

image registration. The image matching method us-

ing phase information called Band-Limited Phase-Only

Correlation (BLPOC) has been proposed57) to dedicate

POC to similarity measure. POC and BLPOC cannot

handle the nonlinear deformation of images, since the

∗ MBGC: https://www.nist.gov/programs-projects/

multiple-biometric-grand-challenge-mbgc∗∗ FOCS: https://www.nist.gov/programs-projects/

face-and-ocular-challenge-series-focs

71


IrisFingerprint Dental X-ray imagePalmprint

Palmprintverification app

Iris verificationunit

2D/3D faceverification unit

Disaster victime identification software:Dental Finder

3D face2D faceFinger knuckle

Biometric recognition using phase-only correlation

Door security Fingerprint verification uint

Fig. 6 Application of phase-only correlation to biometric recognition.

phase information includes only translational displace-

ment. The approach combined with phase-based corre-

spondence matching58) and BLPOC has been proposed

to deal with nonlinear deformation59). So far, we have

applied POC techniques to various biometric recogni-

tion problems59) 63) as shown in Fig. 6. We summarize

(i) the importance of phase information in images, (ii)

fundamentals of POC, BLPOC, correspondence match-

ing and local phase features and (iii) applications to

some biometric recognition problems in palmprint, fin-

ger knuckle and dental in the following.

5. 1 The Importance of Phase Information in

Images

The importance of the phase information in images

has been reported in some literature64)65). Oppenheim65) said that many of the important features of a sig-

nal are preserved if only the phase is retained. We

demonstrate the importance of phase information in im-

ages by replacing phase components between images as

shown in Fig. 7 (Similar discussion has been given in

Refs.65)66)). First, we calculate DFT of Image A and Im-

age B, and obtain amplitude and phase components of

each image. Next, we synthesize new frequency compo-

nents of the image by replacing the phase components

of Image A with those of Image B. Then, we calcu-

late Inverse DFT (IDFT) of the synthesized frequency

components and obtain the new images whose phase

components are replaced. As observed in Fig. 7, the

synthesized images are similar to the image having the

corresponding phase components. This result indicates

that the phase components contain the most important

information to construct the image.

5. 2 Phase-Only Correlation (POC)

As mentioned above, the phase components include

the important information of the image. Accurate im-

age matching can be realized when using only the phase

components. The following describes the fundamental

DFT

DFT

IDFT

IDFT

Amplitude

Phase

Amplitude

Phase

Original image Synthesized image

Image B

Image A

Image B’

Image A’

Fig. 7 The importance of phase information in images

by replacing phase between images: Image A’

with phase components of Image B looks like

Image B. Similarly, Image B’ with phase compo-

nents of Image A looks like Image A. Therefore,

the phase components contain the most impor-

tant information to construct the image.

of POC64)67).

Consider twoN1×N2 images, f(n1, n2) and g(n1, n2),

where we assume that the index ranges are n1 =

−M1, · · · ,M1 (M1 > 0) and n2 = −M2, · · · ,M2 (M2 >

0) for mathematical simplicity, and hence N1 = 2M1+1

and N2 = 2M2 +1. The discussion could be easily gen-

eralized to non-negative index ranges with power-of-two

image size. Let F (k1, k2) and G(k1, k2) denote the 2D

DFTs of f(n1, n2) and g(n1, n2), respectively. Accord-

ing to the definition of DFT68), F (k1, k2) and G(k1, k2)

are given by

F (k1, k2) =∑n1,n2

f(n1, n2)Wk1n1

N1W k2n2

N2

= AF (k1, k2)ejθF (k1,k2), (1)

G(k1, k2) =∑n1,n2

g(n1, n2)Wk1n1

N1W k2n2

N2

= AG(k1, k2)ejθG(k1,k2), (2)

respectively, where k1 = −M1, · · · ,M1, k2 =

−M2, · · · ,M2, WN1= e−j 2π

N1 , WN2= e−j 2π

N2 , and∑n1,n2

denotes∑M1

n1=−M1

∑M2

n2=−M2. AF (k1, k2) and

AG(k1, k2) are amplitude, and θF (k1, k2) and θG(k1, k2)


72

are phase. The normalized cross power spectrum

RFG(k1, k2) is given by

RFG(k1, k2) =F (k1, k2)G(k1, k2)∣∣∣F (k1, k2)G(k1, k2)

∣∣∣= ejθ(k1,k2), (3)

where G(k1, k2) is the complex conjugate of G(k1, k2)

and θ(k1, k2) denotes the phase difference θF (k1, k2) −θG(k1, k2). The POC function rfg(n1, n2) is the 2D

IDFT of RFG(k1, k2) and is given by

rfg(n1, n2) =1

N1N2

∑k1,k2

RFG(k1, k2)

×W−k1n1

N1W−k2n2

N2, (4)

where∑

k1,k2denotes

∑M1

k1=−M1

∑M2

k2=−M2. When two

images are similar, their POC function gives a distinct

sharp peak. When two images are not similar, the peak

drops significantly. The height of the peak gives a good

similarity measure for image matching, and the loca-

tion of the peak shows the translational displacement

between the images.

We have proposed a high-accuracy translational dis-

placement estimation method, which employs (i) an an-

alytical function fitting technique to estimate the sub-

pixel position of the correlation peak, (ii) a windowing

technique to eliminate the effect of periodicity in 2D

DFT, and (iii) a spectrum weighting technique to re-

duce the effect of aliasing and noise67).

5. 3 Band-Limited POC (BLPOC)

We have proposed a BLPOC function dedicated to

the similarity measurement task57). The idea to im-

prove the matching performance is to eliminate mean-

ingless high-frequency components in the calculation of

normalized cross power spectrum RFG depending on

the inherent frequency components of images. Assume

that the ranges of the inherent frequency band are given

by k1 = −K1, · · · ,K1 and k2 = −K2, · · · ,K2, where

0<=K1<=M1 and 0<=K2<=M2. Thus, the effective size

of frequency spectrum is given by L1 = 2K1 + 1 and

L2 = 2K2 + 1. The BLPOC function is given by

rK1K2

fg (n1, n2) =1

L1L2

∑k1,k2

′RFG(k1, k2)

×W−k1n1

L1W−k2n2

L2, (5)

where n1 = −K1, · · · ,K1, n2 = −K2, · · · ,K2, and∑′k1,k2

denotes∑K1

k1=−K1

∑K2

k2=−K2. Note that the

maximum value of the correlation peak of the BLPOC

function is always normalized to 1 and does not depend

on L1 and L2.

I1p1 = p0 / 2

I2 J2

J1

q1 = p1 + 2 δglobal

I0 J0

p0 = ( p0,1, p0,2)

BLPOC

BLPOC

δglobal

δlocal

BLPOC

α

w1

w2

w1

w2

Global registration

Local registration

Similarity evaluationq0 = 2 (q1 + δlocal)

Fig. 8 Flow of hierarchical local block matching using

BLPOC with 3 layers.

5. 4 Phase-based correspondence matching

In order to handle the nonlinear deformation of im-

ages, we employ the approach of correspondence match-

ing using POC58), which employs (i) a coarse-to-fine

strategy using image pyramids for robust correspon-

dence search and (ii) a translational displacement es-

timation method using POC for local block matching.

Let p be a coordinate vector of a reference pixel in the

reference image I(n1, n2). The problem of correspon-

dence search is to find a real-number coordinate vector

q in the input image J(n1, n2) that corresponds to the

reference pixel p in I(n1, n2). Figure 8 shows a flow of

BLPOC-based correspondence matching for biometric

recognition. We briefly explain the procedure as fol-

lows.

Step 1: For l = 1, 2, · · · , lmax, create the l-th layer

images Il(n1, n2) and Jl(n1, n2), i.e., coarser versions

of I0(n1, n2) (= I(n1, n2)) and J0(n1, n2) (= J(n1, n2)),

recursively as follows:

Il(n1, n2) =1

4

1∑i1=0

1∑i2=0

Il−1(2n1 + i1, 2n2 + i2),(6)

Jl(n1, n2) =1

4

1∑i1=0

1∑i2=0

Jl−1(2n1 + i1, 2n2 + i2).(7)

Step 2: Estimate the displacement between Ilmax(n1, n2)

and Jlmax(n1, n2) using BLPOC-based image matching.

Let the estimated displacement vector be δlmax.

Step 3: For every layer l = 1, 2, · · · , lmax, calculate the

coordinate pl = (pl,1, pl,2) corresponding to the original

reference point p0 (= p) recursively as follows:

pl =

⌊1

2pl−1

⌋=

(⌊1

2pl−1,1

⌋,

⌊1

2pl−1,2

⌋), (8)

73


Hierarchical image Extracted features

Fig. 9 Feature extraction from a finger knuckle image,

where “•” indicates the reference point.

where �z� denotes the operation to round the element

of z to the nearest integer towards minus infinity.

Step 4: We assume that qlmax= plmax

+ δlmaxin the

coarsest layer. Let l = lmax − 1.

Step 5: From the l-th layer images Il(n1, n2) and

Jl(n1, n2), extract two local image blocks fl(n1, n2) and

gl(n1, n2) with their centers on pl and 2ql+1, respec-

tively. The size of image blocks is Wc ×Wc pixels.

Step 6: Estimate the displacement between fl(n1, n2)

and gl(n1, n2) using BLPOC-based image matching.

Let the estimated displacement vector be δl. The l-th

layer correspondence ql is determined as follows:

ql = 2ql+1 + δl. (9)

Step 7: Decrement the counter by 1 as l ← l − 1 and

repeat from Step 5 to Step 7 while l >= 0.

Step 8: From the original images I0(n1, n2) and

J0(n1, n2), extract two image blocks with their centers

on p0 and q0, respectively. Calculate BLPOC functions

for all the pairs of two image blocks. The matching

score S is evaluated by

S =Nth

Nblock, (10)

where Nth is the number of image block pairs whose

peak value of the BLPOC function is over the thresh-

old and Nblock is the number of image blocks.

5. 5 Local Phase Features

We have proposed local phase features extracted from

each layer of multi-scale image pyramids, which are de-

signed for biometric recognition69). Figure 9 shows an

example of extracting local phase features from a finger

knuckle image and Fig. 10 shows an example of match-

ing local phase features and an input finger knuckle

image. Using the proposed local phase features, we

can align the global translation between images in the

Input image

−10 −5 0 5 10

−10−5

05

10−0.1

0

0.2

0.3

0.4

0.5

0.1

Registered data

BLPOC

BLPOC

BLPOC

Fig. 10 Matching between hierarchical local phase fea-

tures in Fig. 9 and the input finger knuckle

image.

Fig. 11 Screenshot of the prototype smartphone app of

palmprint recognition.

top (or coarsest) layer, align the minute translation be-

tween local block images in the middle layer, and finally

evaluate the similarity between local block images in

the bottom (or original image) layer. The size of local

phase features can also be reduced by phase quantiza-

tion without sacrificing the performance of biometric

recognition.

5. 6 Applications

We describe some our research results in the follow-

ing: (i) palmprint, (ii) finger knuckle and (iii) dental.

( 1 ) Palmprint

Palmprint recognition is one of the good applications

of POC, since a palm includes rich texture information

to be matched. We have considered a practical contact-

less palmprint recognition system70) 72) based on the re-

sult of the excellent performance of POC in palmprint

recognition69).

A palm image can be easily taken by a built-in camera

of smartphones. We have developed a user authentica-

tion app using palm images70)71) as shown in Fig. 11.

The palmprint recognition algorithm combining a set of


74

(a)

(b)

Fig. 12 Example of hand images, where it is difficult

to extract ROI using the conventional meth-

ods: (a) fingers are closed together (b) a hand

is rolled.

simple image processing, which consists of preprocess-

ing and matching steps, is used to effectively utilize the

limited computational resources of smartphones. The

preprocessing step extracts a hand from the input im-

age using skin-color thresholding and region growing,

detects keypoints and extracts an ROI. The matching

step normalizes affine transformation between ROIs ac-

cording to the correspondence between ROIs obtained

by phase-based correspondence matching and then cal-

culates the matching score. Experimental evaluation

using palmprint image databases demonstrates the effi-

cient performance of the proposed algorithm compared

with conventional algorithms.

We have addressed one of challenging issues in palm-

print recognition72). Accurate ROI extraction is indis-

pensable in contactless authentication, since the perfor-

mance of contactless palmprint recognition significantly

depends on the accuracy of ROI extraction. Therefore,

a variety of hand pose changes must be considered to re-

alize reliable and accurate palmprint recognition. The

conventional approaches of ROI extraction35)73)74) as-

sume that all fingers spread and a palm is not rolled,

since these approaches are based on binarized images to

extract a palm region from a hand image. In practical

situations, this assumption is not always satisfied from

our experience. It is trivial for some persons that fin-

gers are closed together when acquiring a hand image.

In such cases, it is difficult to detect valley points be-

tween fingers and to use a finger shape, since fingers in

the binarized hand image are not separated as shown in

Fig. 12. The public palmprint databases such as PolyU

palmprint database and CASIA palmprint database are

also constructed based on the above assumption. Ad-

dressing the above problem and realizing practical con-

tactless palmprint recognition, we proposed an accurate

and robust palm region extraction method72). The pro-

posed method employs the combination of image bi-

narization and edge detection to detect keypoints as

shown in Fig. 13. The use of the combined approach

makes it possible to detect valley points between fin-

gers accurately, even if fingers are closed and a hand

is rolled. Figure 14 shows some examples of ROI ex-

traction from palm images under unconstrained condi-

tions compared with conventional methods proposed by

Zhang et al.33), Han et al.75) and Leng et al.74). Con-

ventional methods do not extract palm regions on the

correct location or extract palm regions with different

size and location. The proposed method extracts palm

regions whose accuracy is comparable with the ground

truth, since the keypoints are accurately detected by

the proposed method.

( 2 ) Finger Knuckle

An outer surface of a finger has three knuckles: a

distal interphalangeal (DIP) joint, a proximal interpha-

langeal (PIP) joint and a metacarpophalangeal (MCP)

joint as mentioned in Sect. 3. 3. We have developed

a practical person authentication system using PIP

joints63)76) and MCP joints77) for door security. Finger

knuckle patterns can be captured by a camera when a

user takes hold of a door handle. This image acquisition

procedure is not intrusive for the user, since this proce-

dure is a trivial action for everyone to open the door.

Hence, the users do not pay attention to the authen-

tication process. Our systems also used the combined

information of the four knuckles to improve the per-

formance of finger knuckle recognition. In the case of

PIP joints, a camera is embedded into a door so as to

face the camera toward PIP joints as shown in Fig. 15

(a). In the case of MCP joints, a camera is attached

on a door handle as shown in Fig. 15 (b). PIP joints

have rich texture, resulting in better recognition accu-

racy than MCP joints, while all the PIP joints are not

always faced toward a camera due to the structure of

a hand63)76). All the MCP joints can be extracted from

the captured image, resulting in more stable than PIP

joints, while nonlinear deformation of MCP joints has

to be addressed to obtain good performance77). The

accuracy of PIP joint recognition is good, although all

the PIP joints are not always extracted from only one

still images76).

75


Binarizingthe image

Extracting the contour of the hand

Detecting keypoints

Establishing the coordinate system

Extracting the central part

Binarized image Edges Contour and keypointsInput image

Fig. 13 Flow of the proposed ROI extraction method for contactless palmprint recognition.

Manual(Ground truth)

Proposed

Leng

Zhang

Han

Fig. 14 Example of ROI extraction from images under unconstrained conditions.

Door handle

Camera

Handle

NIR light Visible light source

Camera

(a) (b)

Fig. 15 Finger knuckle recognition systems for (i) PIP

joints and (ii) MCP joints.

( 3 ) Dental

Person identification using dental information is one

of the most important works in our research activities.

We summarize techniques of victim identification ac-

tually used in the Great East Japan Earthquake and

Tsunami on March 11, 2011. We also present future

prospects of advanced radiograph-based human identi-

fication techniques, which may have a significant im-

pact on reducing the time and improving the reliability

of large-scale disaster victim identification (DVI).

The Great East Japan Earthquake was a magnitude

9.0 undersea megathrust earthquake off the coast of

Japan that occurred on March 11, 2011. The epicen-

ter of the earthquake is approximately 70 km (43 mi)

east of the Oshika Peninsula in Miyagi Prefecture. The

earthquake triggered huge tsunami waves that reached

heights of up to 40.5 m (133 ft) in Miyako, Iwate Pre-

fecture, and which traveled up to 10 km (6 mi) in-

land in the Sendai area. As of April 10, 2013, the

National Police Agency of Japan has confirmed 15,883

deaths and 2,681 people missing across twelve prefec-

tures. The largest number of victims were confirmed

in Miyagi Prefecture, where 9,537 deaths (60% of the

total deaths) and 1,315 people missing. Forensic den-

tistry78)79)80) played a key role of human identification in

the Great East Japan Earthquake and Tsunami. The

authors have contributed to (A) preparation of stan-

dard instruments package for dental identification, (B)

development of dental record matching software Dental


76

Finder, and (C) design and implementation of overall

workflow of dental identification as shown in Fig. 16.

For more details, please refer to the literature81) and

the project web page∗.We briefly describe the procedure of dental record

matching. Dental record matching is done by compar-

ing each tooth status of antemortem (AM) and post-

mortem (PM) dental records. Dentists have given a

detailed description of each tooth status in AM and

PM dental records. Hence, dentists may make differ-

ent observations of tooth status each other due to a

variety of treatment statuses, even if their observations

are essentially the same. Addressing this problem, we

classify the precise tooth statuses into major classes.

Dental Finder employs 5-class expressions of individual

tooth status as shown in Table 3. The 5-class AM or

PM dental records are input into a database of Dental

Finder using the data input interface. Then, Dental

Finder evaluates the similarity between AM and PM

pairs using the following four similarities: (i) the num-

ber of completely matched teeth, (ii) the number of

matched teeth in class 2 or 3, (iii) the number of teeth

with a consistent state transition and (iv) the match-

ing score. The matching score is calculated using the

weight table for all the tooth pairs between AM and

PM, where the weights are optimized using the known

genuine pairs in advance. In addition to the above sim-

ilarities, we introduce the matching priority to Dental

Finder. The matching priority indicates the possibility

that the top-1 pair is a genuine pair, which is defined

by the difference of similarities between top-1 and top-2

pairs. Therefore, the dentists only have to check from

top-1 pairs having high matching priority. The match-

ing results are provided for all the possible combinations

denoted by “full-combination search” or for the selected

tooth denoted by “individual search.” In practice, we

find matching candidates having high matching prior-

ity using full-combination search and then confirm their

detailed matching results using individual search.

To address future crisis and DVI, we have developed a

novel and automated dental radiograph matching sys-

tem that can assist the task of forensic experts. The

system uses a highly accurate image matching tech-

nique, i.e., POC, in order to find corresponding points

between the two X-ray images, correct image distor-

tion and measure their similarity, as illustrated in Fig.

∗ DVI web page (in Japanese): http://www.aoki.ecei.tohoku.ac.

jp/dvi/

Miyagi Prefectural Police Headquarters

CPU: Core i7 3.46GHz 6 cores 12 threads

Memory: 24GB (DDR3)SSD: 256GB SATA (OS)HDD: 2TB SATA (Data)

Server

Possible candidates ofgenuine AM-PM matching pairs

Dental radiograph matching system(to be used)

Dental FinderDental chart matching systemAntemortem→Postmortem searchPostmortem→Antemortemsearch

Antemortem/Postmortem database

Dental radiographs,oral photographsand other documents

Final decision by the police(Facial appearance, clothes, belongings, fingerprints, palmprints, DNA and etc.)

Comparative identification by dentists

Portable X-ray machine(ADX4000)

Waterproof, shockproof and dust-resistant camera(RICOH G700)

Morgue(Information of dead body)

Dead body

Handwrriten dental chart

Dental records

Oral photographs

Dental radiographs

Dental clinic etc. (Information of missing person)

Dental radiographs(Digital scan)

Convert dental treatmentrecord to dental chart

Fig. 16 Proposed workflow of victim identification in

the Great East Japan Earth Quake.

17. We apply the system to a large-scale identifica-

tion problem, where the system is used to find a spe-

cific individual in a whole radiograph database actu-

ally used in a dental clinic during 2005–200882). The

database consists of 4,810 intraoral radiographs from

1,714 subjects. We randomly select 100 subjects (as

imaginary “disaster victim”) who have at least three

different pairs of radiographs, with each pair taken from

the same oral region before and after dental treatment.

The 100 × 3 radiographs taken after treatment are as-

sumed as “postmortem” (PM) images, and are removed

from the original database. Hence, our “antemortem”

(AM) database contains 4,510 images from 1,714 sub-

jects. Our problem is to search the 100 victims within

AM database using three PM images as the identity key

for every victim. We demonstrated that the proposed

system can reduce the number of pairs to be matched

by forensic experts to only 0.7% (= 33/4, 510) when

three PM radiographs are available.

6. Conclusion

This paper has presented a brief introduction of re-

cent advances in biometric recognition, especially in

face, fingerprint and iris recognition. Researchers seek

more difficult problems such as person authentication

under unconstrained conditions and also new biometric

traits to enhance the accuracy and convenience of bio-

metric recognition. We have also presented our activ-

ities of biometric recognition. Our approach employs

the phase information obtained by Discrete Fourier

Transform (DFT) of images. The correlation func-

tion of phase information, called Phase-Only Correla-

tion (POC), gives us both the good similarity measure

for biometric recognition and the translational displace-

77


Table 3 Correspondence table between dental charts obtained from recovered bodies and 5-class expres-

sion of individual tooth status for Dental Finder.

Class Brief treatment status Detailed treatment status

1 Sound tooth, caries, resin filling, etc. Sealant, wedge-shaped detect, temporary splint, C1, C2, C3, resin filling, cement filling,

glass ionomer filling, root canal filling, incisal edge fracture, remaining tooth, etc.

2 Partial restoration (metal) Inlay, onlay, amalgam filling, 4/5 cast crown, 4/5 temporary crown, etc.

3 Full restoration Resin facing cast crown, metal bond crown, facing cast crown, hard resin jacket crown,

post crown, temporary crown, core, etc.

4 C4 and missing Coping, denture, pontic, missing tooth, unerupted tooth, implant, etc.

5 N/A N/A, lost postmortem, partial loss of body, impacted tooth, etc.

Input image 1 Input image 2

Contrast enhancement

Global registration(translation and rocation)

Distortion correction

Common regionsFig. 17 Dental radiograph matching using POC.

ment for image registration. POC has been success-

fully applied to fingerprint, face, iris, palmprint, fin-

ger knuckle and dental recognition. Among them, we

present some interesting results of palmprint recogni-

tion, finger knuckle recognition and dental recognition.

References

1) A. Jain, R. Bolle, and S. Pankanti, Biometrics: Personal Identi-

fication in a Networked Society, Norwell, MA: Kluwer, 1999.

2) A. Jain, P. Flynn, and A. Ross, Handbook of Biometrics,

Springer, 2008.

3) A. Jain, A. Ross, and S. Prabhakar, “An introduction to biomet-

ric recognition,” IEEE Trans. Circuits and Systems for Video

Technology, vol.14, no.1, pp.4–20, Jan. 2004.

4) A.A. Ross, K. Nandakumar, and A.K. Jain, Handbook of Multi-

biometrics, Springer, 2006.

5) P. Viola and M. Jones, “Robust real-time face detection,” Int’l

J. Computer Vision, pp.137–154, May 2004.

6) S. Zafeirioub, C. Zhanga, and Z. Zhang, “A survey on face de-

tection in the wild: Past, present and future,” Computer Vision

and Image Understanding, pp.1–24, Sept. 2015.

7) S. Yang, P. Luo, C. Loy, and X. Tang, “WIDER FACE: A face de-

tection benchmark,” Proc. IEEE Computer Society Conf. Com-

puter Vision and Pattern Recognition, pp.5525–5533, June 2016.

8) T. Cootes, G. Edward, and C. Taylor, “Active appearance mod-

els,” IEEE Trans. Pattern Anal. Machine Intell., vol.23, no.6,

pp.681–685, 2001.

9) V. Blanz and T. Vetter, “Face recognition based on fitting a 3D

morphable model,” IEEE Trans. Pattern Anal. Machine Intell.,

vol.25, no.9, pp.1063–1074, 2003.

10) J. Booth, A. Roussos, S. Zafeiriou, A. Ponniah, and D. Dunaway,

“3d morphable model learnt from 10,000 faces,” Proc. IEEE

Computer Society Conf. Computer Vision and Pattern Recog-

nition, pp.5543–5552, June 2016.

11) S. Li and A. Jain, Handbook of Face Recognition, Springer, 2011.

12) M. Turk and A. Pentland, “Eigenfaces for recognition,” J. Cog-

nitive Neurosci., vol.3, no.1, pp.71–86, 1991.

13) P. Belhumeur, J. Hespanha, and D. Kriegman, “Eigenfaces vs.

Fisherfaces: recognition using class specific linear projection,”

IEEE Trans. Pattern Anal. Machine Intell., vol.19, no.7, pp.711–

720, 1997.

14) S. Watanabe, P. Lambert, C. Kulikowski, and J. Buxton, “Eval-

uation and selection of variables in pattern recognition,” Com-

puter and Information Sciences II (J.T. Tou Ed.), Academic

Press, New York, pp.91–122, 1967.

15) K. Fukui and O. Yamaguchi, “Face recognition using multi-

viewpoint patterns for robot vision,” Proc. 11th Int’l Symp.

Robotics Research, pp.192–201, Oct. 2003.

16) M. Nishiyama, O. Yamaguchi, and K. Fukui, “Face recognition

with the multiple constrained mutual subspace method,” Proc.

Int’l Conf. Audio- and Video-Based Biometric Person Authenti-

cation, pp.71–80, July 2005.

17) T. Ahonen, A. Hadid, and M. Pietikainen, “Face description with

local binary patterns: Application to face recognition,” IEEE

Trans. Pattern Anal. Mach. Intell., vol.28, no.12, pp.2037–2041,

Dec. 2006.

18) M. Pietikainen, A. Hadid, G. Zhao, and T. Ahonen, Computer

Vision Using Local Binary Patterns, Springer, 2011.

19) Y. Taigman, M. Yang, M. Ranzato, and L. Wolf, “DeepFace:

Closing the gap to human-level performance in face verification,”

Proc. IEEE Computer Society Conf. Computer Vision and Pat-

tern Recognition, pp.1701–1708, June 2014.

20) A. Jain, S. Dass, and K. Nandakumar, “Soft biometric traits

for personal recognition systems,” Lecture Notes in Computer

Science (Proc. Int’l Conf. Biometric Authentication), vol.3072,

pp.731–738, July 2004.

21) W. Scheirer, N. Kumar, K. Ricanek, P. Belhumeur, and T. Boult,

“Fusing with context: A Bayesian approach to combining de-

scriptive attributes,” Proc. Int’l Joint Conf. Biometrics, Oct.


78

2011.

22) Y. Zhong, J. Sullivan, and H. Li, “Face attribute prediction using

off-the-shelf CNN features,” Proc. Int’l Conf. Biometrics, June

2016.

23) D. Maltoni, D. Maio, A.K. Jain, and S. Prabhakar, Handbook of

Fingerprint Recognition, Springer, 2003.

24) L. Hong, Y. Wan, and A. Jain, “Fingerprint image enhancement:

Algorithm and performance evaluation,” IEEE Trans. Pattern

Anal. Mach. Intell., vol.20, no.8, pp.777–789, Aug. 1998.

25) A. Jain, L. Hong, and R. Bolle, “On-line fingerprint verification,”

IEEE Trans. Pattern Anal. Mach. Intell., vol.19, no.4, pp.302–

314, April 1997.

26) J. Fierrez-Aguilar, L. Nanni, J. Ortega-Garcia, R. Cappelli, and

D. Maltoni, “Combining multiple matchers for fingerprint ver-

ification: A case study in FVC2004,” Lecture Notes in Com-

puter Science (Proc. Int’l Conf. Image Analysis and Processing),

vol.3617, pp.1035–1042, Sept. 2005.

27) R. Cappelli, M. Ferrara, and D. Maltoni, “Minutia Cylinder-

Code: A new representation and matching technique for finger-

print recognition,” IEEE Trans. Pattern Anal. Machine Intell.,

vol.32, no.12, pp.2128–2141, Dec. 2010.

28) A. Jain and J. Feng, “Latent fingerprint matching,” IEEE Trans.

Pattern Anal. Mach. Intell., vol.33, no.1, pp.88–100, Jan. 2011.

29) J. Feng, J. Zhou, and A. Jain, “Orientation field estimation for la-

tent fingerprint enhancement,” IEEE Trans. Pattern Anal. Mach.

Intell., vol.35, no.4, pp.925–940, April 2013.

30) A. Sankaran, M. Vatsa, and R. Singh, “Latent fingerprint match-

ing: A survey,” IEEE Access, vol.2, pp.982–1004, Aug. 2014.

31) D. Zhang, Palmprint Authentication, Kluwer Academic Publica-

tion, 2004.

32) A. Kong, D. Zhang, and M. Kamel, “A survey of palmprint recog-

nition,” Pattern Recognition, vol.42, no.7, pp.1408–1418, Jan.

2009.

33) D. Zhang, W. Kong, J. You, and M. Wong, “Online palmprint

identification,” IEEE Trans. Pattern Anal. Mach. Intell., vol.25,

no.9, pp.1041–1050, Sept. 2003.

34) Y. Han, T. Tan, Z. Sun, and Y. Hao, “Embedded palmprint

recognition system on mobile devices,” Lecture Notes in Com-

puter Science (ICB2007), vol.4642, pp.1184–1193, Aug. 2007.

35) G.K.O. Michael, T. Connie, and B.J.T. Andrew, “Touch-less

palm print biometrics: Novel design and implementation,” Image

and Vision Computing, vol.26, pp.1551–1560, July 2008.

36) D. Zhang, W. Zuo, and F. Yue, “A comparative study of palm-

print recognition algorithms,” ACM Computing Surveys, vol.44,

no.1, pp.2:1–2:37, Jan. 2012.

37) A. Kumar and Z. Xu, “Can we use second minor finger knuckle

patterns to identify humans?,” Proc. IEEE Computer Soci-

ety Conf. Computer Vision and Pattern Recognition Workshop,

pp.106–112, June 2014.

38) D. Woodard and P. Flynn, “Finger surface as a biometric iden-

tifier,” Computer Vision and Image Understanding, vol.100,

pp.357–384, Dec. 2005.

39) A. Kumar and Y. Zhou, “Personal identification using finger

knuckle orientation features,” Electronics Letters, vol.45, no.20,

pp.1023–1025, Sept. 2009.

40) L. Zhang, L. Zhang, D. Zhang, and H. Zhu, “Online finger-

knuckle-print verification for personal authentication,” Pattern

Recognition, vol.43, pp.2560–2571, July 2010.

41) L. Zhang, L. Zhang, D. Zhang, and H. Zhu, “Ensemble of lo-

cal and global information for finger-knuckle-print recognition,”

Pattern Recognition, vol.44, pp.1990–1998, Sept. 2011.

42) L. Zhang, L. Zhang, D. Zhang, and Z. Guo, “Phase congruency

induced local features for finger-knuckle-print recognition,” Pat-

tern Recognition, vol.45, pp.2522–2531, July 2012.

43) A. Shahin, A. Badawi, and M. Kamel, “Biometric authentication

using fast correlation of near infrared hand vein patterns,” Int’l

Journal of Biological and Medical Sciences, vol.2, no.3, pp.141–

148, Nov. 2007.

44) N. Miura, A. Nagasaka, and T. Miyatake, “Feature extraction

of finger-vein patterns based on repeated line tracking and its

application to personal identification,” Machine Vision and Ap-

plications, vol.15, no.4, pp.194–203, Oct. 2004.

45) K. Bowyer and M. Burge, Handbook of Iris Recognition,

Springer, 2016.

46) J. Daugman, “High confidence visual recognition of persons by

a test of statistical independence,” IEEE Trans. Pattern Anal.

Mach. Intell., vol.15, no.11, pp.1148–1161, Nov. 1993.

47) S. Shah and A. Ross, “Iris segmentation using geodesic active

contours,” IEEE Trans. Information Frensics and Security, vol.4,

no.4, pp.824–836, Dec. 2009.

48) Z. He, T. Tan, Z. Sun, and X. Qiu, “Toward accurate and fast

iris segmentation for iris biometrics,” IEEE Trans. Pattern Anal.

Mach. Intell., vol.31, no.9, pp.1670–1684, Sept. 2009.

49) A. Ross, “Iris recognition: The path forward,” Computer, vol.2,

no.43, pp.30–35, Feb. 2010.

50) K. Ricanek, M. Savvides, D. Woodard, and G. Dozier, “Uncon-

strained biometric identification: Emerging technologies,” Com-

puter, vol.2, no.43, pp.56–62, Feb. 2010.

51) V. Pauca, M. Forkin, X. Xu, R. Plemmons, and A. Ross, “Chal-

lenging ocular image recognition,” Proc. SPIE 8029, Sensing

Technologies for Global Health, Military Medicine, Disaster Re-

sponse, and Environmental Monitoring; and Biometric Technol-

ogy for Human Identification VIII, no.80291V, May 2011.

52) V. Boddeti, J. Smereka, and B. Kumar, “A comparative evalua-

tion of iris and ocular recognition methods on challenging ocular

images,” Proc. Int’l Joint Conf. Biometrics, pp.1–8, Oct. 2011.

53) S. Crihalmeanu and A. Ross, “Multispectral scleral patterns

for ocular biometric recognition,” Pattern Recognition Letters,

vol.14, no.33, pp.1860–1869, Oct. 2012.

54) A. Ross, R. Jillela, J. Smereka, V. Boddeti, B. Kumar,

R. Barnard, X. Hu, P. Pauca, and R. Plemmons, “Matching

highly non-ideal occular images: An information fusion ap-

proach,” Proc. Int’l Conf. Biometrics, pp.446–453, April 2012.

55) L. Xiao, Z. Sun, and T. Tan, “Fusion of iris and periocular bio-

metrics for cross-sensor identification,” Lecture Notes in Com-

puter Sciences (Proc. 7th Chinese Conf. Biometric Recognition),

vol.7701, pp.202–209, Dec. 2012.

56) M. Monwar, B. Vijayakumar, V. Boddeti, and J. Smereka, “Rank

information fusion for challenging ocular image recognition,”

Proc. IEEE Int’l Conf. Cognitive Informatics & Cognitive Com-

puting, pp.175–181, July 2013.

57) K. Ito, H. Nakajima, K. Kobayashi, T. Aoki, and T. Higuchi,

“A fingerprint matching algorithm using phase-only correlation,”

IEICE Trans. Fundamentals, vol.E87-A, no.3, pp.682–691, March

2004.

58) K. Takita, M.A. Muquit, T. Aoki, and T. Higuchi, “A sub-pixel

correspondence search technique for computer vision applica-

tions,” IEICE Trans. Fundamentals, vol.E87-A, no.8, pp.1913–

1923, Aug. 2004.

59) K. Ito, S. Iitsuka, and T. Aoki, “A palmprint recognition algo-

rithm using phase-based correspondence matching,” Proc. Int’l

Conf. Image Processing, pp.1977–1980, Nov. 2009.

60) K. Ito, T. Aoki, T. Hosoi, and K. Kobayashi, “Face recognition

using phase-based correspondence matching,” Proc. IEEE Conf.

Automatic Face and Gesture Recognition, pp.173–178, March

2011.

61) K. Miyazawa, K. Ito, T. Aoki, K. Kobayashi, and H. Nakajima,

“An effective approach for iris recognition using phase-based im-

age matching,” IEEE Trans. Pattern Anal. Mach. Intell., vol.30,

no.10, pp.1741–1756, Oct. 2008.

62) S. Aoyama, K. Ito, and T. Aoki, “A finger-knuckle-print recog-

nition algorithm using phase-based local block matching,” Infor-

mation Sciences, vol.268, pp.53–64, June 2014.

63) D. Kusanagi, S. Aoyama, K. Ito, and T. Aoki, “Multi-finger

knuckle recognition from video sequence: Extracting accurate

multiple finger knuckle regions,” Proc. Int’l Joint Conf. Biomet-

rics, Sept. 2014.

64) C.D. Kuglin and D.C. Hines, “The phase correlation image

alignment method,” Proc. Int’l Conf. Cybernetics and Society,

pp.163–165, 1975.

65) A.V. Oppenheim, “The importance of phase in signals,” Proc.

IEEE, vol.69, no.5, pp.529–541, May 1981.

66) M. Savvides, B.V.K.V. Kumar, and P.K. Khosla, “Eigenphases

vs eigenfaces,” Proc. 17th Int’l Conf. Pattern Recognition, vol.3,

pp.810–813, Aug. 2004.

67) K. Takita, T. Aoki, Y. Sasaki, T. Higuchi, and K. Kobayashi,

“High-accuracy subpixel image registration based on phase-

only correlation,” IEICE Trans. Fundamentals, vol.E86-A, no.8,

pp.1925–1934, Aug. 2003.

68) A. Oppenheim, R. Schafer, and J. Buck, Discrete-Time Signal

Processing (2nd Edition), Prentice Hall, 1999.

69) S. Aoyama, K. Ito, and T. Aoki, “Similarity measure using lo-

cal phase features and its application to biometric recognition,”

Proc. IEEE Computer Society Conf. Computer Vision and Pat-

tern Recognition Workshop, pp.180–187, June 2013.

70) S. Aoyama, K. Ito, T. Aoki, and H. Ota, “A contactless palmprint

79


recognition algorithm for mobile phones,” Proc. Int’l Workshop

on Advanced Image Technology, no.6A-5, pp.409–413, Jan. 2013.

71) H. Ota, S. Aoyama, R. Watanabe, K. Ito, Y. Miyake, and

T. Aoki, “Implementation and evaluation of a remote authentica-

tion system using touchless palmprint recognition,” Multimedia

Systems, vol.19, no.2, pp.117–129, March 2013.

72) K. Ito, T. Sato, S. Aoyama, S. Sakai, S. Yusa, and T. Aoki,

“Palm region extraction for contactless palmprint recognition,”

Proc. Int’l Conf. Biometrics, pp.334–340, May 2015.

73) W. Jia, R. Hu, J. Gui, Y. Zhao, and X. Ren, “Palmprint recog-

nition across different devices,” Sensors, vol.12, no.6, pp.7938–

7964, 2012.

74) L. Leng, G. Liu, and M. Li, “Logical conjunction of triple-

perpendicular-directional translation residual for contactless

palmprint preprocessing,” Proc. Int’l Conf. Information Technol-

ogy: New Generations, pp.523–528, April 2014.

75) H. Han, Z. Sun, F. Wang, and T. Tan, “Palmprint recognition

under unconstrained scenes,” Lecture Notes in Computer Science

(ACCV2007), vol.4844, no.2, pp.1–11, Nov. 2007.

76) S. Aoyama, K. Ito, and T. Aoki, “A multi-finger knuckle recog-

nition system for door handle,” Proc. IEEE Sixth Int’l Conf.

Biometrics: Theory, Applications and Systems, no.O-18, Sept.

2013.

77) D. Kusanagi, S. Aoyama, K. Ito, and T. Aoki, “A practical per-

son authentication system using second minor finger knuckles for

door security,” IPSJ Trans. Computer Vision and Applications,

vol.9, no.8, pp.1–13, March 2017.

78) D. Senn and P. Stimson, Forensic Dentistry, CRC Press, 2010.

79) I. Pretty and D. Sweet, “A look at forensic dentistry — Part 1:

The role of teeth in the determination of human identity,” British

Dental J., vol.190, no.7, pp.359–366, April 2001.

80) D. Sweet and I. Pretty, “A look at forensic dentistry — Part 2:

Teeth as weapons of violence — identification of bitemark per-

petrators,” British Dental J., vol.190, no.8, pp.415–418, April

2001.

81) T. Aoki and K. Ito, “What is the role of universities in disaster

response, recovery, and rehabilitation? Focusing on our disaster

victim identification project,” IEEE Comm. Magazine, vol.52,

no.3, pp.30–37, March 2014.

82) E. Kosuge, K. Ito, Y. Hanzawa, and T. Aoki, “Large-scale per-

formance evaluation of a dental radiograph matching system

for forensic human identification,” Proc. Radiological Society of

North America 2009, pp.1069–1070, Nov. 2009.

Koichi Ito received the B.E. degree in elec-tronic engineering, and the M.S. and Ph.D. de-gree in information sciences from Tohoku Univer-sity, Sendai, Japan, in 2000, 2002 and 2005, respec-tively. He is currently an Assistant Professor of theGraduate School of Information Sciences at TohokuUniversity. From 2004 to 2005, he was a ResearchFellow of the Japan Society for the Promotion ofScience. His research interest includes signal andimage processing, and biometric authentication.

Takafumi Aoki received the BE, ME, andDE degrees in electronic engineering from TohokuUniversity, Sendai, Japan, in 1988, 1990, and 1992,respectively. He is currently a professor in theGraduate School of Information Sciences (GSIS)at Tohoku University. Since April 2012, he hasalso served as the Vice President of Tohoku Uni-versity. His research interests include theoreticalaspects of computation, computer design and orga-nization, LSI systems for embedded applications,digital signal processing, computer vision, imageprocessing, biometric authentication, and securityissues in computer systems. He received more than20 academic awards as well as distinguished serviceawards for his contributions to victim identificationin the 2011 Great East Japan Disaster.


80

Date post:	10-Dec-2021
Category:	Documents
Upload:	others
View:	3 times
Download:	0 times

Paper Recent Advances in Biometric Recognition

Documents