www.ietdl.

org
Published in IET Biometrics
Received on 16th August 2012
Revised on 8th March 2013
Accepted on 8th April 2013
doi: 10.1049/iet-bmt.2012.0047

ISSN 2047-4938

Dental radiographs and photographs in human

forensic identification
Vijayakumari Pushparaj, Ulaganathan Gurunathan, Banumathi Arumugam
Department of ECE, Thiagarajar College of Engineering, Madurai, Tamil Nadu, India
E-mail: vijayakumari@tce.edu

Abstract: Dentistry can contribute for the identification of human remains after any disasters or crimes in assistance to other
medical specialties. The algorithm can be developed by comparing post mortem and ante mortem dental radiographs. This
work aims to introduce photographic images in addition to radiographs. In this research a contour and skeleton-based shape
extraction as well as matching algorithm for dental images is proposed. An active contour model with selective binary and
Gaussian filtering regularised level set method is used for contour extraction. Shape matching is done by both contour and
skeleton-based approaches. The experimental results are obtained from a database of dental images include both radiographs
and photographs. This algorithm provides better matching decision about the person than the existing algorithms since it
includes skeleton measures also. The performance measures obtained and the hit-rate indicates that the better matching is
observed with radiographic than the photographic images.

1 Introduction present in multiple digitised dental records in order to

Forensic odontology is a branch of forensics that deals with
victim identification based on dental features. Owing to the
evolution of information technology and an urge to
investigate more cases by the forensic experts, it is
necessary to automate the human identification system.
Dental pattern can be considered as a biometric if there is
no other means of physiological biometrics such as palm
print, finger print, iris, face, leg print etc., are found.
Human identification using dental images has been proven
to be the best under certain circumstances if there is
inaccessibility of other means of biometrics. Teeth and
bones are treated as the hardest and robust tissues found in
the human body, which yields resistant to modest force
effects, high temperatures up to 1100°C [1]. It also
possesses good biometric properties. It affords resistance to
decay even in some critical situations as well as mass
disasters such as bomb-blast, air-crash, major fire accident
or flood etc. In a recent disaster with a count rate of
thousands of victims, around 50–70% of the cases are
identified using dental records, 20–35% using fingerprints
and 3–20% using DNA evidence, respectively [2]. The
dental evidence is considered as the most suited biometric
for martyr identification in the Indian Ocean earthquake,
resultant tsunami on 26th December 2004 and Thailand
tsunami attack on January 2005 [3]. Automating the
postmortem identification of deceased individuals based on
dental characteristics is receiving increased attention
especially with the large number of victims encountered.
This initiates the usage of dental images as a better choice
for individual and mass disaster victims identification. An
automated dental identification system compares the teeth
access their similarity. In some situations, if the dental
radiographs are unavailable or severely distorted, the
identification can be done by photographic images also. In
case of both radiographs and photographs, the algorithm
requires both shape extraction and matching. There are
several approaches for dental radiograph segmentation. Jain
and Chen [4], were dedicated a concept of semi-automatic
contour method for shape extraction and pattern matching.
The shortcomings in their approach include, if the image is
too blurred and severely occluded, their algorithm may not
be pertinent. The computation time of this algorithm is
higher since crown and root shape extraction had been done
separately. For contour extraction the morphological corner
detection produces comparatively better hit-rate whereas it
also fails to handle severely occluded dental radiographs
[5]. The shape extraction is further efficient by using
connected and fast connected component labelling [6],
which uses Mahalanobis distance as the measure of
matching. The person matching was analysed [7] with
various similarity and distance metrics. Said et al. [8]
offered a mathematical morphology approach, which uses a
series of morphology filtering operations to improve the
segmentation, and by using connected component analysis
the desired region of interests are obtained. Human
identification is also explained using shape and appearance
of the tooth [9], in which panoramic dental images are not
handled. Nomir and Abdel-Mottaleb [10] achieved the
efficiency improvement by combining three different
matching techniques. Although performing three different
levels of matching is computationally complex, it seems to
be efficient. Matching of dental records using hierarchical
distance [11] proves to be an efficient in terms of retrieval

56 IET Biom., 2013, Vol. 2, Iss. 2, pp. 56–63

& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-bmt.2012.0047
 www.ietdl.org
time. Based on analyses of tooth anatomy and tooth growth
direction, a dental radiograph segmentation algorithm was
developed [12]. Individual identification is supported by
classification and numbering of teeth. Mesiodistal neck
detection is introduced [13] for molar and premolar
classification. One of the notable issues of automated
dental identification system is the missing tooth. A
concept of finding missing tooth using classification and
numbering was done in order to aid content-based
retrieval of dental images [14]. It is differently dealt
with multi slice computed tomography images using
multi-resolution wavelet-Fourier descriptors for classifying
the teeth sequence which is invariant to geometrical
transformation [15]. Individual person identification may
not be perfect while using anatomy of teeth alone. In case
of absence or inaccuracy of dental records, the person
identification can be done by family photographs [16]. It
is evident from Thailand tsunami victim identification
that, weak and absence of dental records did not stop
forensic odontology team from their contribution towards
person identification. It can even be done with the
availability of photograph of upper anterior teeth [17]. In
case of inadequate availability of dental radiographs, it
could be interesting to analyse the family albums or
photographs taken during some functions for the missing
person identification based on its dimension, size and
alignment of teeth [18]. Active contour model or snakes,
proposed by Kass et al. [19], has been proved to be an
efficient framework for image segmentation. The Fig. 1 Pipe line of the proposed method
fundamental idea of active contour model is to start with
a curve around the object to be detected, and the curve
moves towards its interior normal and stops on the true The dental radiographs with missing tooth is one of the
boundary of the object based on an energy-minimising research issues addressed earlier. Consequently, this research
model. Level set method is based on active contour work exploit radiographic images and photographic images
model and particularly designed to handle the with missing tooth, dental work such as crown mineralisation
segmentation of deformable structures. Generally, the and filling etc. This research will explain about the
classical active contour model uses spline curves to model skeleton-based measures in addition with the contour-based
the boundary of an object. However, the level set method approach both for the dental radiographs and photographs.
is to use a deformable curve front for approximating the The paper work is organised as four sections. The pipeline of
boundary of an object. In the level set framework, the this approach is shown in Fig. 1. The first section is contour
curve is represented by the zero level set of a smooth tracing using selective binary Gaussian filtering and
function, usually called the level set function. Moving the regularised level set (SBGFRLS) method. In the second
curves can be done by evolving the level set functions section, shape extraction is done by skeleton. The third
instead of directly moving the curves. Therefore level set section is shape matching of the contours traced. Observing
methods exhibit interesting elastic behaviours and can shape matching using skeleton is the fourth section.
efficiently handle the topological changes which is also a
main advantage compared with classical active contour 2 Shape extraction
model. In general, the goal of most active contour
algorithms is to extract the boundaries of homogeneous For individual identification, exact shape information is
regions within an image, whereas the goal of most needed. It is done here in two ways. One method is
anisotropic diffusion algorithms is to smooth the values of contour-based and the other one is skeleton based.
an image within homogeneous regions but not across the
boundaries of such regions. Simultaneous image 2.1 SBGFRLS method
smoothing and segmentation algorithm was developed
[20] using curve evolution algorithm. The initial contour In this section, a region-based active contour model with a
selection and the interior contour growing concepts are SBGFRLS method is introduced. It uses a new
presented using active contours without edges [21]. For region-based SPF function, which can efficiently stop the
some of the images the algorithm fails to produce exact contours at weak or blurred edges. The advantage of this
contour. Chan and Vese solved the minimisation problem technique is automatically detecting the interior and exterior
using level set method following Mumford–Shah [22]. boundaries with the initial contour being anywhere in the
Mathematical relationship between the general image. It is possible to segment not only the desired object
formulations of parametric and geometric active contours but also the other objects. By combining the geodesic
was detailed [23]. Following Chan–Vese model signed active contour and Chan–Vese contours, a SPF function is
pressure force (SPF) function is used as energy function designed [24]. It modulates the signs of the pressure forces
for active contours [24]. The computational complexity of inside and outside the region of interest so that the contour
Chan–Vese model is reduced modestly in this algorithm. shrinks when outside the object or expands when inside the

IET Biom., 2013, Vol. 2, Iss. 2, pp. 56–63 57

doi: 10.1049/iet-bmt.2012.0047 & The Institution of Engineering and Technology 2013
www.ietdl.org
object. The signal flow graph of SBGFRLS method is shown Here ∇ is the gradient operator and α is the balloon force,
in Fig. 2. The SPF function is which controls the contour shrinking or expanding. div


  ((∇φ)/(|∇φ|)) is the curvature-based term used to regularise
I(x) − c1 + c2 /2 the level set function. The traditional level set algorithm
SPF(I(x)) =      (1) possesses some problems. It includes using an signed
max I(x) − c1 + c2 /2  distance function instead of SPF and it also requires
re-initialisation. After every iteration, a Gaussian filter is
where I(x) is the region of interest of the image; c1 and c2 are used to regularise the selective binary level set function in
the two constants which are the average intensities inside and order to overcome the problems in the traditional existing
outside the contour, respectively. In the Chan–Vese model, methods. The simplified level set formulation using
the constants c1 and c2 are computed using level set method as Gaussian filter is written as

I(x)H(f) dx ∂f    
c1 (f) = V
(2) = spf I(x) a∇f, x[V (6)
V
H(f) dx ∂t
  
I(x) 1 − H(f) dx The level set function φ can be initialised as
V  
c2 (f) = (3) ⎧
1 − H(f) dx
V ⎨ −r; x [ V0 − ∂V0
f(x, t) = 0 = 0; x [ ∂V0 (7)
where H(φ) is the Heaviside function. Since the signs of the ⎩
SPF function inside and outside the object are opposite, it r ; x [ V − V0
can be written as
where ρ > 0 is a constant, Ω0 is a subset in the image domain
c + c2   Ω and ∂Ω0 is the boundary of Ω0. The selective step is helpful
Min(I(x)) , 1 , Max I(x) , x[V (4) to select the desired object by considering φ = 1 if φ > 0;
2
otherwise φ = −1. The selection of parameter Gσ, standard
deviation of Gaussian filter is a notable one here. If it is too
Here, consider Ω be a bounded open subset of R 2 and I: [0, small, it will be sensitive to noise and if it is too large, the
a] × [0, b] → R + be a given image. Let C (q): [0, 1] → R 2 be edge leakage may occur and will cause inaccurate
a parameterised planar curve in Ω. By using (1), the level boundary. For dental images it is observed in the range of
set formulation is derived as 25–50.
 
∂f   ∇f  
= spf I(x) div   + a ∇f 2.2 Skeletal representation
∂t ∇f
  Skeleton is a nice shape descriptor because skeleton matching
+ ∇spf I(x) ∇f, x [ V (5) can perform as part matching, that is, part of the whole image
may be used as the region for matching. Another advantage of
skeleton matching is that, it is an intuitive representation of
shape and can be understood by the user, allowing the user
to be more controlled in the matching process. Skeleton is
obtained using iterative thinning operation. Thinning is a
morphological operation, which is used to remove some
selected foreground pixels from the binary image. It is
commonly used to tidy up the output of edge detectors by
reducing all lines to single pixel thickness.

3 Shape matching
As like shape extraction, shape matching is the next desirable
factor. Shape matching is one of the fundamental problems in
computer vision. In this work, it is achieved in two ways. It is
observed with both distance measures and then skeletal
measures to make obvious decision about victim
identification.

3.1 Contour matching with distance metric

The ante-mortem dental records might have captured a long

before than the post mortem images. Hence the viewing
angle might be differing in both the dental records. So,
there is a necessity of applying a rigid transformation to
both ante mortem and post mortem images before finding
the distance. The matching distance obtained can be
improved by the rigid transformation. It is of the form

Fig. 2 Flow graph of the SBGFRLS method x′ = Rx + t (8)

58 IET Biom., 2013, Vol. 2, Iss. 2, pp. 56–63

& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-bmt.2012.0047
 www.ietdl.org
where x′ is the transformation on x, R is the transformation images better will be the matching performance. The
matrix and t is the translation vector. R and t are radiograph includes periapical and bitewing images. Some
represented as of the images are taken with dental work and missing
  tooth cases. The sample input images are shown in Fig. 3.
cos u sin u Sx 0 The size of the test images Dent19, Dent12, Dent25,
R= (9)
−sin u
cos u 0 Sy Dent16, tooth7, photo91 and photo92 are 159 × 160, 74 ×
98, 198 × 150, 157 × 160 144 × 103, 199 × 109 and 160 ×

t 115, respectively. The contour and skeletons obtained for
t= x (10) the radiographic image Dent16 was shown in Fig. 4.
ty
Fig. 4a is the maxilla and mandible regions of Dent16.
Fig. 4b is its corresponding contour. Their equivalent
Here, θ is the rotation angle, tx and ty are the horizontal and
skeleton representation is shown in Fig. 4c. To perform
vertical translation parameters, Sx and Sy are the horizontal
matching in a convenient way, maxilla and mandible
and vertical scaling parameters. The difference in both the
regions are shown separately. The consequent outputs for
ante-mortem and post-mortem images can be observed by
photographic images are shown in Fig. 5. Original
using the Euclidean distance (ED). It is given as
photographic images taken are exposed in Fig. 5a. Since

n  2 upper jaw alone will be predominantly seen for most of
ED = i=1
X i − Yi (11) the persons when they laugh, both the images shown here
is focused with maxilla only. Fig. 5b is the grey equivalent
where Xi and Yi refers to each and every pixel in both the of the input. The contours traced for these images are
images. If the matching distance is low, better matching shown along with the input in Fig. 5c. For convenience the
may be expected. contour alone is extracted out of it and shown in Fig. 5d.
The contour trace was obtained using active contour with
SBGFRLS method and the skeleton is obtained using
3.2 Skeletal-based matching iterative thinning process. With these contour and skeleton
information, matching of both ante-mortem and
In conventional methods, computing a set of skeletal nodes,
post-mortem dental records can be performed. The
connecting the nodes into a graph and then indexing into a
contours thus extracted are helpful in missing tooth
database and verification with one or more objects is
followed. In this paper additional parameters are considered
for skeleton matching. They are (i) the centroid, (ii)
distance of the skeleton end point from the origin (iii)
length of the skeleton and (iv) angle of skeleton end point
with respect to the reference point. The x and y coordinates
of the centroid is denoted xc, yc and is obtained as

1 N
−1
1 N
−1
xc = x; yc = y (12)
N i=0 i N i=0 i

where N is the number of points on the boundary. The ED and

the length can be obtained using (11). The angle is measured
between the two arms. By considering the upper end point of Fig. 3 Sample input images taken
the skeleton as one arm (a), and the gap valley point (b) as the a Radiographic images
other, the angle θi is defined as b Photographic images


yb − ya
ui = arctan (13)
xb − xa

Here, x and y are the co-ordinates of (a) and (b).

4 Results and discussion

The algorithm is implemented with MATLAB R2010b,
using a Pentium IV CPU (3.00 GHz-Dual Core) on a
Microsoft Windows XP environment. This algorithm was
evaluated with a marginal database of 62 dental
radiographs include images captured from either the left or
the right jaw and a limited database of 42 dental
photographic images captured in the front view. In addition
with this, the database of 300 individual teeth with
maximum of molar, premolar and a minimum of canine Fig. 4 Contour and skeletons obtained for the radiographic image
teeth are formed. In the same way, for photographic Dent16
analysis the database of 150 individual teeth which a Input radiographic images
includes a maximum of central, lateral incisors and a b Contours traced using SBGFRLS method
minimum of canine teeth are formed. Higher the database c Equivalent skeleton representation

IET Biom., 2013, Vol. 2, Iss. 2, pp. 56–63 59

doi: 10.1049/iet-bmt.2012.0047 & The Institution of Engineering and Technology 2013
www.ietdl.org
radiographic images, whereas the traditional level
set algorithms require more iteration to converge and takes
around 45 min [24]. Similarly it converges within 100
iterations and takes only 0.1 min for photographic images.
It is a bit computationally simpler and faster than
radiographic images. This algorithm accurately detects the
exterior and interior boundaries of the dental images taken.
The initial contour is selected far from the region of interest
of the image. For making the matching process still
efficient, the skeleton of all the images are obtained as
shown in the figures.
For individual identification, instead of having the whole
image as such, the single tooth or two teeth are yielding
better matching. It is more suitable in case of photographic
images, because the middle two teeth in the upper jaw
alone can be considered for 70% of matching. The resultant
images for both radiograph and photograph with single and
two teeth are shown in Figs. 7 and 8, respectively.
To obtain perfect matching, distance measure alone may
not be a sufficient one. Accordingly, the proposed method
uses few more measures using skeleton-based matching. It
includes the length of the skeleton from the origin, centroid
point of the skeleton, the angle between the skeleton of a
particular tooth from the reference point (gap valley) and
the ED of each skeleton from the origin. These measures
are obtained for seven radiographic images and two

Fig. 5 Consequent outputs for photographic images

a Original photographic image
b Corresponding grey image
c Contours grown on the actual image
d Extracted contour
e Equivalent skeleton

identification also. If there is no curve grown or gap in the

skeleton image indicates the absence of teeth as in Figs. 5c
and d. It is also notable from Fig. 5c is that the lipstick
applied on the lips is also mistakenly picked up as a
contour. It can be overcome by converting the image into
binary rather than grey as shown in Fig. 6.
In this approach, the evolution of the level set algorithm
converges in 125 iterations and takes only 0.2 min for

Fig. 7 Contours and skeletons for radiographic single and two

teeth cases

Fig. 8 Contours and skeleton for photographic single and two

Fig. 6 Contour extracted with binary image teeth cases

60 IET Biom., 2013, Vol. 2, Iss. 2, pp. 56–63

& The Institution of Engineering and Technology 2013 doi: 10.1049/iet-bmt.2012.0047
 www.ietdl.org
Table 1 Skeletal measures Table 2 Matching distance
Images Measures Input images taken Matching distance, pixels

Length, Angle, ED, Centroid, Ante-mortem Post-mortem Contour based Skeleton based
pixels deg. pixels co-ordinates
single4 single7 8.07 7.824
Dent1 30 78.41 20 (31.7, 27.1) single5 6.843 6.23
tooth4 29 147.66 83 (29.3, 88.8) single3 8.483 7.679
Dent5 41 134.6 47 (30.1, 54.9) tooth4 8.892 8.245
Dent17 41 111.08 83 (84.3, 28.6) photo91 Ph1 6.28 5.524
Dent20 169 82.35 68 (76.3, 53.5) Pho3 16.23 12.578
single4 41.9 128.46 83 (31.7, 37.3) photo93 21.45 18.67
single5 41.9 131.1 83 (37.6, 33.0) photo92 5.36 5.21
photo92 32.1 86.4 22 (21.7, 31)
photo91 34 88.2 22 (23.9, 32.1)

The precision of a reference image Y and the test image X is

defined as
photographic images such as Dent1, Dent5, tooth4, Dental17,
Dent20, single4, single5, photo92 and photo91 are tabulated |Y > X |
P(X , Y ) = (14)
in Table 1. The skeletal measures for few of the images are |X |
also point out in Fig. 9. From the Table 1 it is evident that,
for the images Dent1 and tooth4, although the length is Recall can be calculated by
closer other measures are conflicting a lot. Likewise, Dent5
and Dent17 are having slightly closer measures for length |Y > X |
R(X , Y ) = (15)
and angle, the other measures are contradictory; whereas, |Y |
single4 and single5 are having exact matching of length,
ED and other two measures are also closely related. Hence,
better matching can be expected for this case. The same
two images are seemed to have lesser matching distance for
both the approaches as shown in Table 2. The similar concept
is evident for photographic images also. Photo91 and
photo92 are having better skeletal measures and minimum
matching distance. Thus, matching is decided in terms of
both these measures. Another inference from Table 2 is that,
skeleton-based approach is yielding comparatively minimum
distance with contour-based approach.

4.1 Performance evaluation

The algorithm is evaluated by observing precision, recall,
fall-out, F-measure and overall. These measures are
obtained for both of these approaches and are plotted in
Figs. 10 and 11.

Fig. 10 Performance measures of contour approach

Fig. 9 Skeleton measures of radiographic and photographic

images
a Binary image of single tooth
b Centroid representation
c Skeleton
d Skeleton with angle measure Fig. 11 Performance measures of skeleton approach

IET Biom., 2013, Vol. 2, Iss. 2, pp. 56–63 61

doi: 10.1049/iet-bmt.2012.0047 & The Institution of Engineering and Technology 2013
www.ietdl.org
The percentage of false positive retrieved pairs can be
observed by using fallout measure
| X | − |X > Y |
F(X , Y ) = (16)
|X |
F-measure is used to aggregate the results of precision and
recall. It is obtained as
P(X , Y ) · R(X , Y )
Ma (X , Y ) = (17)
(1 − a) · P(X , Y ) + a · R(X , Y )
Here α is assumed to be 0.5, which gives the harmonic mean
of precision and recall. The overall measure is obtained as
 for photographic images also with fewer computations. The
1
O(X , Y ) = R(X , Y ) × 2 − (18)
P(X , Y )
These measures are obtained for three pairs of database and
query images with both contour and skeleton. From
Figs. 10 and 11, it is clear that recall and precision
parameters are having inverse relations and overall value is
lesser than F-measure. Fallout parameter is lesser in both.
Although comparing contour and skeleton-based approaches
the precision and overall values are higher in skeleton than
contour-based approach. This algorithm is also evaluated
based on the cumulative matching characteristic curve
(CMC). It is plotted by having the values of top ranking
percentage and the hit-rate percentage of trials. It is
revealed in Fig. 12. From the characteristic curve plotted, it
is observed that a high hit-rate of 0.77 is obtained for
radiographic images with both contour and skeleton
approaches. It is slightly higher than [4, 5]; whereas a
slightly lesser hit-rate of 0.69, 0.67 is obtained for
photographic images with contour and skeleton approaches,
respectively. For radiographic images with contour
approach, 87% of images are retrieved within top-3
rankings and 98.3% retrieval is observed within top-15
rankings. Similarly, 83.8%, 96.7% is retrieved within top-3,
top-15 rankings correspondingly for radiographic skeleton
approach. The photographic analysis yields 80.9%, 78.5%
retrieval only within top-3 rankings for contour and
skeleton, respectively.
Fig. 12 Cumulative matching characteristics
62
& The Institution of Engineering and Technology 2013
5 Conclusion
Developing an automated dental identification system is a
demanding challenge at present. In this paper a novel shape
matching algorithm using skeleton is proposed for dental
images. Another novel focus of this paper is usage of dental
photographs if there is unavailability of dental radiographs. It
is an attempt to provide an aid for forensic law enforcement
with the help of photographic images also. The contour
tracing is implemented using a level set method named
SBGFRLS method. This contour tracing algorithm holds
good even for bitewing images with dental works. Since
matching with contour alone may not produce convincing
results, an additional information using skeleton is worn in
this paper. The experimental results clearly show that the
algorithm which is adapted to radiographic images is suited
precision and overall measures are higher for skeleton than
contour, while considering the whole image, either maxilla
or mandible separately for matching. Whereas for individual
tooth, radiographic images with contour based approach is
better than photographic images. In future, some more
additional descriptors can be considered for photographic
images in order to improve the performance.
6 Acknowledgment
We are very much grateful to Dr. M. Rajkumar, Professor and
Head, Forensic Department, Tanjore Medical College,
Tanjore for his valuable suggestions and the Department of
ECE, Thiagarajar College of Engineering, Madurai, Tamil
Nadu for providing all the facilities to carry out this work.
7 References
1 Savio, C., Merlati, G., Danesino, P., Fassina, G., Menghini, P.:
‘Radiographic evaluation of teeth subjected to high temperatures:
experimental study to aid identification processes’, Forensic Sci. Int.,
2006, 158, (2), pp. 108–116
2 Auerkari, E.: ‘Recent trends in dental forensics’, Indonesian J. Legal
Forensic Sci., 2008, 1, (1), pp. 5–12
3 Petjua, M., Suteerayongprasertb, A., Thongpudc, R., Hassirid, K.:
‘Importance of dental records for victim identification following the
Indian Ocean tsunami disaster in Thailand’, J. R. Inst. Public Health,
2007, 121, pp. 251–257
4 Anil Jain, K., Chen, H.: ‘Matching of dental X-ray images for human
identification’, Pattern Recognit., 2004, 37, pp. 1519–1532
5 Banumathi, A., Vijayakumari, B., Raju, S.: ‘Performance analysis of
various techniques applied in Human identification using Dental
X-Rays’, J. Med. Syst., 2007, 31, (3), pp. 210–218
6 Pushparaj, V., Gurunathan, U., Arumugam, B.: ‘An effective dental
shape extraction algorithm using contour information and matching by
mahalanobis distance’, J. Digit. Imaging, 2013, 26, (2), pp. 259–263,
doi 10.1007/s10278–012-9492-4
7 Pushparaj, V., Gurunathan, U., Arumugam, B.: ‘Human forensic
identification using similarity and distance metrics’. Proc. Eighth
Annual IEEE India Conf. INDICON, Cochin, India, December 2012
8 Said, E.H., Nassar, D.E., Gamal Fahmy, M., Hany Ammar, H.: ‘Teeth
segmentation in digitized dental X-ray films using mathematical
morphology’, IEEE Trans. Inf. Forensics Sec., 2006, 1, (2), pp. 178–189
9 Nomir, O., Abdel-Mottaleb, M.: ‘Human identification from Dental
X-Ray images based on the shape and appearance of the teeth’, IEEE
Trans. Inf. Forensics Sec., 2007, 2, (2), pp. 188–197
10 Nomir, O., Abdel-Mottaleb, M.: ‘Fusion of matching algorithms for
human identification using dental X-ray radiographs’, IEEE Trans. Inf.
Forensics Sec., 2008, 3, (2), pp. 223–233
11 Nomir, O., Abdel-Mottaleb, M.: ‘Hierarchical contour matching for
dental X-ray radiographs’, Pattern Recognit., 2008, 41, (1), pp. 130–138
12 Hosntalab, M., Zoroofi, R.A., Tehrani-Fard, A.A., Shirani, G.:
‘Automated dental recognition in MSCT images for human
identification’. Fifth Int. Conf. Intelligent Information Hiding and
Multimedia Signal Processing-IEEE, 2009, pp. 1318–1321
IET Biom., 2013, Vol. 2, Iss. 2, pp. 56–63
doi: 10.1049/iet-bmt.2012.0047

13 Lin, P.L., Lai, Y.H., Huang, P.W.: ‘An effective classification and
numbering system for dental bitewing radiographs using teeth region
and contour information’, Pattern Recognit., 2009, 45, (3), pp. 1–13
14 Aeini, F., Mahmoudi, F.: ‘Classification and numbering of posterior
teeth in bitewing dental images’. Third Int. Conf. Advanced Computer
Theory and Engineering (ICACTE), 2010, pp. 66–72
15 Hosntalab, M., Zoroofi, R.A., Tehrani-Fard, A.A., Shirani, G.:
‘Classification and numbering of teeth in multi-slice CT images using
wavelet-Fourier descriptor’, Int. J. CARS, 2010, 5, pp. 237–249
16 Tinoco, R.L.R., Martins, E.C., Daruge Jr. E., Prado, F.B., Caria, b.P.H.
F.: ‘Dental anomalies and their value in human identification: A case
report’, J. Forensic Odontostomatol., 2010, 28, (1), pp. 39–43
17 Al-Amad, S.H.: ‘Forensic odontology’, Smile Dental J., 2009, 4, (1),
pp. 22–24
18 Silva, R.F., Pereira, S.D., Prado, F.B., Daruge Jr. E.: ‘Forensic
odontology identification using smile photograph analysis – case
reports’, J. Forensic Odontontostomatol., 2008, 27, (1), pp. 12–17
IET Biom., 2013, Vol. 2, Iss. 2, pp. 56–63
doi: 10.1049/iet-bmt.2012.0047
www.ietdl.org
19 Kass, M., Witkin, A., Terzopoulos, D.: ‘Snakes: active contour models’,
Int. J. Comput. Vis., 1988, 1, pp. 321–331
20 Tsai, A., Yezzi, A., Allan Willsky, S.: ‘Curve evolution implementation
of the mumford–shah functional for image segmentation, denoising,
interpolation, and magnification’, IEEE Trans. Image Process., 2001,
10, (8), pp. 1169–1186
21 Tony Chan, F., Luminita Vese, A.: ‘Active contours without edges’,
IEEE Trans. Image Process., 2001, 10, (2), pp. 266–277
22 Lie, J., Lysaker, M., Tai, X.C.: ‘A binary level set model and some
application to Munford–Shah image segmentation’, IEEE Trans.
Image Process., 2006, 15, pp. 1171–1181
23 Xu, C.Y., Yezzi Jr. A., Prince, J.L.: ‘On the relationship between
parametric and geometric active contours’. Processing of 34th
Asilomar Conf. Signals Systems and Computers, 2000, pp. 483–489
24 Zhang, K., Zhang, L., Song, H., Zhou, W.: ‘Active contours with
selective local or global segmentation: A new formulation and level
set method’, Image Vis. Comput., 2010, 28, pp. 668–676
63
& The Institution of Engineering and Technology 2013