Towards Automated Pose Invariant 3D Dental Biometrics Xin ZHONG1, Deping YU1, Kelvin W C FOONG2, Terence SIM3, Yoke San WONG1 and Ho-lun CHENG3 1. Mechanical Engineering, National University of Singapore, 117576,
[email protected] 2. Faculty of Dentistry, National University of Singapore, 119083 3. School of Computing, National University of Singapore, 117417
conditions and mass disasters because DNA is fragile that its structure is easily altered or destroyed through time, heat, chemical or other forces. Traditionally, the identification based on dental radiograph comparisons is labor-intensive and low in efficiency. There are several computer-aided postmortem (PM) identification systems, such as the famous CAPMI [3] and WinID [4]. However, these systems are text-based searching of records and do not provide high level of automation as the feature extraction, coding, and image comparison are still carried out manually. Extensive efforts have been put into the research towards automated two-dimensional (2D) radiograph-based dental identification in the last decade. The 2D framework mainly involves four steps [5]: image segmentation [6], feature extraction [6, 7], atlas registration [8, 9] and matching [10, 11].
Abstract A novel pose invariant 3D dental biometrics framework is proposed for human identification by matching dental plasters in this paper. Using 3D overcomes a number of key problems that plague 2D methods. As best as we can tell, our study is the first attempt at 3D dental biometrics. It includes a multi-scale feature extraction algorithm for extracting pose invariant feature points and a triplet-correspondence algorithm for pose estimation. Preliminary experimental result achieves 100% rank-1 accuracy by matching 7 postmortem (PM) samples against 100 ante-mortem (AM) samples. In addition, towards a fully automated 3D dental identification testing, the accuracy achieves 71.4% at rank-1 accuracy and 100% at rank-4 accuracy. Comparing with the existing algorithms, the feature point extraction algorithm and the triplet-correspondence algorithm are faster and more robust for pose estimation. In addition, the retrieval time for a single subject has been significantly reduced. Furthermore, we discover that the investigated dental features are discriminative and useful for identification. The high accuracy, fast retrieval speed and the facilitated identification process suggest that the developed 3D framework is more suitable for practical use in dental biometrics applications in the future. Finally, the limitations and future research directions are discussed.
(a) (b) Fig. 1 Dental plasters (a) an AM madibular plaster of a live person (b) a PM mandibular plaster of a dry skull
However, many unsolved problems and challenges limit the identification capability and accuracy of the 2D methodology, including: 1) radiographs are often blurred images, making it very difficult to extract the tooth contours accurately with minimal geometric distortions. Moreover this process is often time-consuming. Chen et al. [12] reported that 14 of the 25 subjects in their database could not be identified due to poor image quality , variation of the dental structure and insufficient number of AM images for matching. 2) 2D radiographs are projections of 3D teeth. Distortions in tooth shape arising from different imaging angles are often significant, which causes incorrect matching, namely tooth contours extracted from genuine samples (paired PM and AM samples of a victim) could not be matched together. In contrast, 3D dental identification based on the digitized dental plaster is able to overcome the
1. Introduction Dental biometrics utilizes dental features for victim identification. The use of teeth in postmortem (PM) identification has gained increasing attention over the last half-century. In forensic dentistry, the postmortem identification of a deceased individual is based on the dental records when other evidences of the victim (e.g. clothing, jewelry, pocket contents, gender, estimated age, height, build, color of skin, scars, moles, tattoos, abnormalities, DNA, fingerprints, iris etc.) are not available [1]. Due to the survivability and diversity of dental features, identification by dental records outperforms that by DNA [2] in severe
978-1-4577-1359-0/11/$26.00 ©2011 IEEE
1
AM Dentition
AM Plaster Casts
AM Digitized Model
Decimation
PM Dentition
PM Plaster Casts
PM Digitized Model
Decimation
Feature Point Detection Feature Point Detection
Correspondence(algorithm comparison Experiment III and IV)
Fine Matching
AUTO PCA-plane Segmentation Manual or AUTO PCA-plane Segmentation(Experim ent I and II)
Matching Score
Rank List
Fig.2 An overview of 3D dental biometrics framework aforementioned limitations because 1) laser-scanned 3D dental plasters are high-resolution surface data; 2) projection from 3D to 2D is not required, thus no distortion of the tooth shape occurs. The problem arising from different imaging angles in 3D is what we call pose variation problem which we aim to solve in this paper. Trends in 3D dental biometrics There has been much interest and development in the investigation of 3D dental biometrics. With the development of real-time scanning and 3D reconstruction technologies from 2D images or video sequences, the acquisition of 3D models has become effortless and fast. 3D biometrics is receiving increasingly more attention than 2D biometrics. For instance, 3D face and ear recognition [13, 14] showed a promising future. In addition, there are some emerging dental research works in assisting 3D reconstruction of teeth from CT images [15] and 3D automatic teeth segmentation for dental biometrics [16]. Therefore, the present study aims to investigate 3D identification scheme in dental biometrics by matching dental plasters, such as the two shown in Fig. 1. Our paper makes the following contributions: 1. We propose a novel 3D pose invariant dental biometrics framework. As best as we can tell, ours is the first attempt at 3D dental biometrics; all existing works use only 2D images. It overcomes a number of key hurdles in traditional 2D methods, thus making our method more useful. 2. Our method is fast and could be fully automatic and thus can be used for rapid identification of large groups of people. It takes about 1.7 hours to retrieve one subject from 33 subjects and 7 hours to retrieve from 133 subjects (PC with a 2.99 GHz Pentium 4 processor) [17]. In contrast, it takes only 25 minutes on average to retrieve one subject from 100 subjects. (PC with 2 Duo CPU 2.33 GHz 1.96GB RAM).
3. Our method is faster and more robust to pose variations, which is shown in Experiment III and IV. 4. The dental arch (the curving structure formed by the teeth in their normal position), tooth crown shape and the arrangement of teeth (teeth neighboring position) are used directly without projection to 2D in our study. We discover that the discriminability of these dental features is useful and distinguishable enough to provide potential identities among individuals without tedious single tooth segmentation and contour extraction requirements. In addition, our method is more robust because we can use the dental arch for identification even when individual teeth have been damaged.
2. System Approach An overview of the 3D dental biometrics framework is shown in Fig. 2. Ante-mortem (AM) database The AM database comprises 100 mandibular teeth samples scanned using Minolta VIVID 900 Surface Laser Scanner (Konica-Minolta Corporation, Osaka, Japan). Postmortem (PM) database The PM samples used to match with the AM samples consist of 7 plasters of mandibular teeth which are separately prepared and scanned by a different investigator using the same scanner without knowing the previous scanning parameters. The initial orientations are seldom the same when genuine samples are prepared and scanned by different investigators. In addition, it is observed that even the appearances of the genuine samples are different as can be seen in Fig. 3. The PM sample in Fig. 3 (b) looks smooth compared with its AM sample (Fig. 3 (a)), e.g. some holes are presented in the AM sample. The reason of these differences could be 1) the physical dental plasters are made
2
by different investigators; 2) the different resolutions of scanning; and 3) handling errors during scanning. To facilitate efficient and accurate matching of corresponding AM and PM samples, preprocessing of the digitized samples is required to reduce the size of the sample. The preprocessing comprises three operations: 1) decimation for both AM and PM samples; 2) PCA-plane segmentation for 100 AM samples; and 3) manual/ auto PCA-plane segmentation for 7 PM samples (two experiments). Decimation for both AM and PM samples Each digitized sample is 14~30MB, comprising of 340k~400k triangles and so is decimated by 90% to achieve higher computational speed. Only 10% of the original mesh is used for identification in our present study. We want to show that a competitive accuracy can be achieved by using our proposed approaches even after such large-scale decimation. The decimation algorithm in [18] was utilized. The decimated samples are shown in Fig. 3.
shown in Fig. 4(a). A dental plaster was segmented by its PCA-plane into the crown part and bottom part as illustrated in Fig.4 (b) and Fig.4(c) respectively. Manual/auto segmentation for 7 PM samples. The gum and teeth for the PM samples are to be exactly segmented. It is manually performed because segmentation for a PM sample, which still contains tooth gum, is different from that for the mandibular teeth of a human skull as shown in Fig. 5(a). Most of the 3D segmentation methods detect the interstice between gum and teeth (gingival margin) by computing the points located at minimum curvatures on meshes. If this minima rule is applied to madibular teeth of a skull as shown in Fig.5 (a), the dash line in Fig. 5 (b) will be detected which is the interstices between the teeth and alveolar bone, instead of the expected solid line which is the real interstice between teeth and gum (gingival margin) as shown in Fig. 5(b). Thus one portion of the tooth root, which does not exist in its corresponding AM plaster sample, will be included in PM sample. It will produce error in the matching process. According to forensic dentists’ experience, gum begins to decay within two or three days after death. Therefore, it is quite common to see PM samples without gums. Based on the aforementioned reasons, manual segmentation is implemented to segment madibular teeth of skulls according to the gingival margin. The segmented teeth are shown in Fig. 5(c). In addition, we also test fully automatic identification process in experiment II by applying the same PCA-plane segmentation method to the 7 PM samples in experiment II.
(a) (b) Fig.3 Difference in genuine samples after decimation (a) AM sample of victim I (b) PM sample of victim I
PCA-plane segmentation for 100 AM samples. For a large AM database, the automatic segmentation is necessary. As best as we know, no fully automatic 3D segmentation method achieves a promising accuracy. This is the most tedious and time-consuming step both in 3D and 2D dental biometrics. Some researchers are working towards this goal in orthodontics planning studies. Kondo et al. [19] proposed a highly automatic tooth segmentation method. The dental arch is used to calculate the panoramic range image. However, four reference points need to be manually specified by users at the beginning. Kronfeld et al. [20] presented a highly automatic segmentation method for separation of teeth from the mesh model by applying an active contour algorithm. However, they reported that manual adjustment is still needed when the initial snakes are not appropriately located at the transition between teeth and gum. Both methods fail where the boundary between tooth and gum is very smooth or in severe malocclusion cases. In this study, instead of single-tooth segmentation, a fast automatic processing method is proposed for a large AM database to eliminate the bottom part of the plaster which does not contain tooth information. The Principal Component Analysis (PCA)-plane passing through the centroid of the plaster was calculated for each AM plaster as
(a) (b) (c) Fig. 4 PCA-plane segmentation for an AM sample (a) PCA-plane (b) segmented tooth crown (c) bottom part of a dental plaster
(a) (b) (c) Fig. 5 Manual segmentation of a human skull (a) a human skull (b) the expected detected interstices (solid line) and the interstices obtained by minima curvature rule (dash line) (c) a set of manual segmented mandibular teeth of a human skull
Feature point detection. The principle of key feature point or salient feature point detection is well-established in 2D image processing [21, 22]. During the last decade, several studies have extended it to the 3D domain [23-25]. Inspired by these studies, a multi-scale feature point detection algorithm is presented to extract feature points on digitized
3
dental surfaces. Fig.6 shows the differences between the existing work [23-25] and this work. The main steps are given below. The first step of the feature point detection is computing multi-scale representations for dental mesh surface by applying N Gaussian filters on it. For each vertex v in the surface model, the neighborhood N (v, ) is point x i within distance . As the Euclidean distance gives better results than the geodesic distance[23], equation
N (v, ) x x v , x : vertex
Later, we compare the number of extracted points and the total time in matching genuine samples and imposter samples. It is about six times faster using feature point detection algorithm in this work in matching one PM sample to its genuine AM sample. The results are shown in Experiment III in the next section.
(1)
is used for calculating the neighborhood points. A representation of the surface model G(v, ) can be obtained using equation 2 xi exp xi v / (2 2 ) x N ( v ,2 )
(a) (b) Fig. 6 Feature points on dental meshes (a) existing work (b) this work
G ( v, )
i
2 exp xi v / (2 2 ) x N ( v ,2 )
.
Correspondence Let P’ and Q’ be the feature points extracted from the PM dental surface and the AM dental surface respectively. For each feature point pi P ' and
(2)
qi Q ' , the respective saliency value S ( pi ) and S ( qi ) were already calculated in the feature point detection stage. The following triplet-constrain algorithm is presented to find the best transformations. This step is to find three feature points both in PM and AM samples with similar saliency values and similar relative positions in Euclidean space for correspondence. For any feature point p P ' , select the salient points q as potential correspondence if S ( p) S (q) , where ε is threshold value and set
i
The second step of feature point detection is saliency map computation of dental mesh surfaces. To compute the mesh saliency, the Difference-of-Gaussian (DoG) for each vertex v is defined: (3) DoG (v) G (v, i ) G (v, k i ) as the difference between its Gaussian-weighted representation at scale ( i ) and scale ( k i ). DoG( v ) is actually a 3D vector which denotes the displacement between different scales. Six scales were used σi {1ε, 2ε, 3ε, 4ε, 5ε, 6ε }, where ε is 0.3% of the length of the diagonal of the bounding box of the dental surface model. In order to promote the small number of distinctive high peaks while suppressing the large number of similar high peaks in the saliency map, each saliency map is normalized using the non-linear suppression operator S proposed by Itti et al [21]. The third step is boundary effect removal. The following algorithm is applied: 1) search for the boundary vertices 2) search for the vertices within distance 2σ6 to the boundary vertices; 3) set the saliency of all these vertices to zero. The fourth step is feature point extraction. The saliency map at each scale is processed such that each saliency value is set to zero unless it is larger than the saliency of 85% of its neighboring vertices. The final saliency map for the surface model is then obtained by adding the saliency map at all six scales followed by a normalization process. Finally, a vertex whose saliency value is a local maximum and larger than 60% of the global maximum is detected as a salient point. As shown in Fig. 6(a), edge points are detected as feature points by the existing work [23-25]. Usually, more feature points require more computational time in finding correspondence at the next stage. The edge points are not feature points of tooth shape. The feature points detected by this work with edge effect removal are shown in Fig. 6(b).
to be 0.1 in our tests. Therefore, a set of potential correspondences for each feature point are determined and designated as (C(p1), …, C(pn)). For each pair of feature points (pi, pj), choose any qi C ( pi ) , q j C ( p j ) and set the point pair (qi, qj) which minimizes the distance root mean squared (dRMS) error defined in equation dRMS 2 ( P ', Q ')
1 n
2
n
n
( i 1 j 1
pi p j qi q j )2
(4)
as the associated correspondence pair, resulting in a set E2 of two-point correspondences. E2 is then sorted in order of ascending dRMS error. Any e E 2 whose dRMS error is larger than a threshold dRMS is discarded. For each two-point correspondence e E 2 , add another potential correspondence pair (pk, qk) which minimizes the dRMS error. In this way, a set E3 of triplet-point correspondence is formed. E3 is then sorted in order of ascending dRMS error. Any e E 3 whose dRMS error is larger than a threshold dRMS is discarded. For each triplet-point correspondence in E3, a rotation and translation matrix can be obtained by Singular Value
4
Decomposition (SVD) method and the corresponding coordinate root mean square (cRMS) error is then computed using equation cRMS 2 ( P, Q) min R ,t
1 n Rpi t qi n i 1
2
.
samples are also expected. In our preliminary study, it is interesting to investigate the identification accuracy if all the process are automated. Therefore, two experiements are designed. Experiment I Identification process with human interaction in PM segmentation Experimental results show fully correct priority ranking accuracy based on matching of 7 manually segmented PM samples to a database of 100 AM samples. At rank 1, 100% accuracy was achieved. The retrieval performance curve, as shown in Fig. 9, is often used to evaluate the accuracy of the experiment. The x axis represents the rank of retrieved subjects. Identification of 7 PM samples from 100 AM samples, each PM sample has 100 possible ranks. The y axis indicates the cumulated number of correct retrievals at each rank. Experiment II Fully automated identification process without human interaction In Experiment I, the PM segmentation is the only manual part of the whole identification process. In Experiment II, the same 7 PM samples are segmented using the same PCA-plane segmentation method for AM samples, namely a portion of gum has not been exactly segmented and attached to the teeth. Undoubtedly, the gum and plaster portion will bring errors but the identification process becomes fully automated. We try to test identification accuracy under a rough segmentation condition. The results are shown in Fig. 9. Five out of seven achieved rank-1 accuracy(5/7=71.4%); 6 out of 7 achieved rank-2 accuracy (6/7=85.7%) and at rank 4, 100% accuracy was achieved. Experiment III Feature point extraction algorithms comparison We compare the number of extracted points and the computational time in matching two samples between the existing algorithm and this work by using the same computer. The initial positions of the two samples are the same. We test both genuine samples and imposter samples. The results are shown in Table 1. All the calculations in this paper include time (second) for model importing, visualization. By using this work, the computational total time for matching one pair samples is reduced to 1/6~1/5 (139/22=6.3;151/29=5.2). Experiment VI Correspondence algorithms comparison regarding pose invariant characteristic We compare the similar existing work greedy algorithm [26] with this work. We show that ours is more robust to pose variations. The results are shown in Table 2. The rotation variation is designed to simulate the possible real rotations in scanning. There is a base plane (almost a parallel plane to the principal plane we calculated in Fig.4) the plaster is placed on this plane with the teeth side facing the scanner. Therefore, most rotation variation is around the normal to this plane. Subsequently, we increased 30 degree every time until 360 degree rotation. And we also test the imposter samples. Results show that this work always gave the correct
(5)
Finally, E3 is sorted in order of cRMS error. The first triplet-point correspondence in E3 corresponding to minimal cRMS error is taken as the best triplet-point correspondence. Fig. 7 shows the correspondence in genuine samples. We compare the existing work[26] with this work in Experiment IV in the next section. We show more robust characteristics of this work regarding pose invariant.
Fig. 7 Triplet-point correspondence in genuine samples Fine Matching With the estimated initial position by feature points correspondence, the fine comparisons are achieved by utilizing iterative closest point (ICP) algorithm which was first developed by Besel and Mckay [27], Chen and Medioni[28]. The results of genuine matching and imposter matching of samples in Fig.8. The comparison shows that genuine samples require less iterations and the matching error is much smaller.
(a)
(b) Fig. 8 Fine matching of samples in Fig 8. (a) genuine samples (b) imposter samples
3. Experimental Results Towards an automatic 3D dental identification system development, an automatic segmentation method for PM
5
matching while the existing work [26] failed in most cases. The reason is the previous work is developed for general shapes, such as animal shapes which have visual salient points at ear tips, mouths, claws, nose tips, failing in corresponding dental mesh with highly similar convex and concave, saddle points. We have run further experiments to show that our triplet-constraint algorithm is indeed robust: we injected destructive noises, and our algorithm was still able to correctly locate the corresponding points, even when significant noise was added. Due to page limitation, we are unable to give further details.
This work
Rotation 60 degree Existing work [26] This work
Rotation 90 degree Existing work[26]
This work
Rotation 180 degree Existing work [26]
Fig.9 Comparisons of identification accuracy between a user-intervention process (Experiment I) and a fully automated process (Experiment II)
Table 1 Feature extraction algorithms comparison (Experiment III) Number of Total time points (second) Genuine Existing AM I 118 139 samples work PM I 103 [23-25] This work AM I 48 22 PM I 30 Imposter Existing AM II 139 151 samples work PM I 103 [23-25] This work AM II 74 29 PM I 30
This work
4. Conclusions and Future Work A novel pose invariant 3D dental biometrics framework has been proposed in this paper. As best as we know, our work is the first attempt at 3D dental biometrics; all existing works use only 2D images. A feature point extraction algorithm and a triplet-correspondence algorithm are developed for pose estimation of dental meshes. Experimental results show that the developed algorithms are faster and more robust than the existing ones for pose estimation. We also facilitate the identification process by using 3D dental features directly, avoiding tedious single tooth segmentation and contour extraction processes. We discover that the discriminability of these dental features is enough to provide potential identities. 100% rank-1 accuracy is achieved with user interaction in segmentation by retrieving 7 subjects from 100 subjects. In addition,
Table 2-Experiment VI Correspondence algorithms comparison Rotation 30 degree Existing work[26]
6
[12] H. Chen, A.K. Jain, Dental Biometrics: Alignment and Matching of Dental Radiographs, IEEE Transactions on Pattern Analysis and Machine Intelligence, 27 (2005) 1319-1326. [13] L. Xiaoguang, A.K. Jain, Deformation Modeling for Robust 3D Face Matching, Pattern Analysis and Machine Intelligence, IEEE Transactions on, 30 (2008) 1346-1357. [14] C. Hui, B. Bhanu, Efficient Recognition of Highly Similar 3D Objects in Range Images, Pattern Analysis and Machine Intelligence, IEEE Transactions on, 31 (2009) 172-179. [15] S. Tohnak, A.J.H. Mehnert, M. Mahoney, S. Crozier, Synthesizing Dental Radiographs for Human Identification, J. Dent. Res., 86 (2007) 1057-1062. [16] D. Mairaj, S.D. Wolthusen, C. Busch, Teeth Segmentation and Feature Extraction for Odontological Biometrics, in: Intelligent Information Hiding and Multimedia Signal Processing (IIH-MSP), 2010 Sixth International Conference on, 2010, pp. 323-328. [17] H. Chen, AUTOMATIC FORENSIC IDENTIFICATION BASED ON DENTAL RADIOGRAPHS, in: Department of Computer Science and Engineering, PhD Thesis, Michigan State University, 2007. [18] W.J. Schroeder, J.A. Zarge, W.E. Lorensen, Decimation of triangle meshes, SIGGRAPH Comput. Graph., 26 (1992) 65-70. [19] T. Kondo, S.H. Ong, K.W.C. Foong, Tooth segmentation of dental study models using range images, Medical Imaging, IEEE Transactions on, 23 (2004) 350-362. [20] T. Kronfeld, D. Brunner, G. Brunnett, Snake-based segmentation of teeth from virtual dental casts, Computer-Aided Design and Applications, 7 (2010) 221-233. [21] L. Itti, C. Koch, E. Niebur, A model of saliency-based visual attention for rapid scene analysis, IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 20 (1998) 1254-1259. [22] D.G. Lowe, Distinctive image features from scale-invariant keypoints, International Journal of Computer Vision, 60 (2004) 91-110. [23] C. Lee, A. Varshney, D. Jacobs, Mesh saliency, in: SIGGRAPH '05: ACM SIGGRAPH 2005 Papers, ACM, 2005, pp. 659-666. [24] Y.-S. Liu, M. Liu, D. Kihara, K. Ramani, Salient critical points for meshes, in: SPM 2007: ACM Symposium on Solid and Physical Modeling, June 4, 2007 - June 6, 2007, Association for Computing Machinery, Beijing, China, 2007, pp. 277-282. [25] U. Castellani, M. Cristani, S. Fantoni, V. Murino, Sparse points matching by combining 3D mesh saliency with statistical descriptors, in, Blackwell Publishing Ltd, 9600 Garsington Road, Oxford, OX4 2XG, United Kingdom, 2008, pp. 643-652. [26] N. Gelfand, N.J. Mitra, L.J. Guibas, H. Pottmann, Robust global registration, in: Proceedings of the third Eurographics symposium on Geometry processing, Eurographics Association, Vienna, Austria, 2005, pp. 197. [27] P.J. Besl, A Method for Registration of 3-D Shapes, IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, 14, NO. 2 (1992) 239-256. [28] M. Chen, Object modeling by registration of multiple range images, 1991 IEEE International Conference on Robotics and Automation, 1991. Proceedings., , vol.3 (1992) 2724 - 2729 [29] CJIS division-ADIS, digitized radiographic images (database), (August 2002).
71.4% rank-1 accuracy is achieved in fully automated identification process. The single subject retrieval time has also been significant reduced compare to that using 2D identification framework. There is no 3D dental biometrics benchmark database and the 2D database is not publicly available[29]. Although the comparisons to 2D are not based on the same dataset, our preliminary work is to provide a new vision into dental biometrics by using 3D identification framework which aims to overcome limitations in previous 2D work while facilitating the whole identification process. The retrieval efficiency, accuracy and capability have shown the feasibility of the proposed 3D framework. However, there are some limitations. The data used in this work are dental plasters which only contain the tooth crown shapes. Thus the tooth root and dental work (tooth fillings) are not available which are also useful for dental identification. The samples size is still small. Therefore, our future work could include 1) sample acquisition from Computed Tomography (CT) or Magnetic Resonance Imaging (MRI) images; 2) further testing on a larger database; and 2) other efficient geometric invariant features extraction and correspondence algorithms development.
5. References [1] D.R. Senn, P.G. Stimson, Forensic Dentistry, Second Edition ed., CRC Press, Taylor& Francis Group, 2010. [2] Dental records beat DNA in tsunami IDs, New Scientists, 2516:12 (2005). [3] R.M. Lorton L, Frideman R, The computer-assisted postmortem identification (CAPMI) system: a computer-based identification program., Journal of Forensic Science, (1988) 997-984. [4] M. J, WinID3 dental identification system, in, 2006. [5] H. Chen, A.K. Jain, Automatic Forensic Dental Identification Handbook of Biometrics (2008) 231-251. [6] A.K. Jain, H. Chen, Matching of dental X-ray images for human identification, Pattern Recognition, 37 (2004) 1519-1532. [7] H. Chen, A.K. Jain, Tooth contour extraction for matching dental radiographs, in: ICPR 2004. Proceedings of the 17th International Conference on Pattern Recognition, 2004, pp. 522-525 [8] M.H. Mahoor, M. Abdel-Mottaleb, Automatic classification of teeth in bitewing dental images, in: ICIP '04. International Conference on Image Processing, 2004, pp. 3475-3478 [9] P.L. Lin, Y.H. Lai, P.W. Huang, An effective classification and numbering system for dental bitewing radiographs using teeth region and contour information, Pattern Recognition, 43 (2010) 1380-1392. [10] O. Nomir, M. Abdel-Mottaleb, Fusion of matching algorithms for human identification using dental X-ray radiographs, IEEE Transactions on Information Forensics and Security, 3 (2008) 223-233. [11] O. Nomir, M. Abdel-Mottaleb, Hierarchical contour matching for dental X-ray radiographs, Pattern Recognition, 41 (2008) 130-138.
7
