Your SlideShare is downloading. ×
  • Like
FutureTech 2010
Upcoming SlideShare
Loading in...5
×

Thanks for flagging this SlideShare!

Oops! An error has occurred.

×

Now you can save presentations on your phone or tablet

Available for both IPhone and Android

Text the download link to your phone

Standard text messaging rates apply

FutureTech 2010

  • 91 views
Published

 

Published in Technology , Education
  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
    Be the first to comment
    Be the first to like this
No Downloads

Views

Total Views
91
On SlideShare
0
From Embeds
0
Number of Embeds
0

Actions

Shares
Downloads
1
Comments
0
Likes
0

Embeds 0

No embeds

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

Cancel
    No notes for slide

Transcript

  • 1. May 20 - 24, 2010 | Busan, Korea 1 “An Efficient Ear Identification System” Authors: D. R. Kisku*, S. Gupta, P. Gupta, J. K. Sing *Department of Computer Science and Engineering, Dr. B. C. Roy Engineering College, Durgapur – 713206, West Bengal, India Email: drkisku@ieee.org
  • 2. May 20 - 24, 2010 | Busan, Korea 2 Agenda: • Biometrics system: Physiological and behavioral characteristics • Introduction to Ear biometrics • Advantages of ear biometrics • State-of-the-art ear biometrics systems • Ear modeling using GMM • K-L divergence for color similarity measurements • SIFT Keypoints Extraction • Fusion strategy of keypoints – Fusion using concatenation approach – Fusion using Dempster-Shafer theory • Experimental results • Conclusion • Bibliography
  • 3. May 20 - 24, 2010 | Busan, Korea 3 Biometric Systems: Physiological and Behavioral Characteristics • Biometric systems – Biometric systems uniquely recognizing humans based upon one or more intrinsic physical or behavioral characteristics. Biometric systems are primarily used in identity access management and access control security. • Biometric characteristics can be divided into two groups – Physiological characteristics are related to the shape of the body. Examples include fingerprint, face, hand and palm geometry, ear, DNA, iris, retina and odor/scent. – Behavioral characteristics are related to the behavior of a person. Examples include gait, voice and typing rhythm.
  • 4. May 20 - 24, 2010 | Busan, Korea 4 Introduction to Ear biometrics: • Ear biometrics have been used as major feature in forensic science for many years. • Ear evidence found in many crime scenes have been used as proofs for hundreds cases in Netherland and United Sates for many years. • Human ear contains large number of unique features and even the ear shape allows for human identification and verification task. • Ear images can be taken from a distance without being knowledge of examined person or subject. • It can be successfully used in access control, identity management and surveillance systems.
  • 5. May 20 - 24, 2010 | Busan, Korea 5 Sample Ear Image from IIT Kanpur Database Easy ear image Complex ear image
  • 6. May 20 - 24, 2010 | Busan, Korea 6 Advantages of Ear Biometrics: • Ear shape does not change over time and ageing • Robust to lighting conditions • Robust to spatial distributions of pixels • Robust to non-uniform distributions of intensity • Deals with background clutter, occlusions, pose variations, etc.
  • 7. May 20 - 24, 2010 | Busan, Korea 7 State-of-the-art Ear Biometrics Systems: • Ear recognition system based on…. – force field transformation [5] – PCA based [6] and LDA-kernel [2] – Feature and geometric measurements [4-5]. – Block-based multi-resolution techniques using wavelet transform and Local Binary Pattern (LBP) [10]. – planer surface and creating a homography transform using SIFT features which lead to be ears being made registered accurately. It can also dealt with background clutter, occlusion and pose variations. – feature level fusion approach [11].
  • 8. May 20 - 24, 2010 | Busan, Korea 8 Proposed Ear Identification System: • Gaussian mixture model [13] is used for modeling the skin color of the ear image. • K-L divergence algorithm [14] is then used to cluster the whole ear image into a number of color slice regions by recording the color similarity properties from a pair of ear images. • From the clustered ear image, SIFT keypoint features [8] are extracted from each grayscale slice region. • To improve the robustness and performance of the system, two techniques, viz. concatenation and Dempster-Shafer decision theory, are used to fuse the invariant features extracted from the slice regions. • Finally, authenticity has been established by using two distance measures, viz. Euclidean distance [12] and Nearest Neighbor approaches [12].
  • 9. May 20 - 24, 2010 | Busan, Korea 9 Ear Modeling using GMM: • An ear image is considered as a collection of coherent regions. • Each homogeneous color region is represented by a Gaussian distribution in the image plane and Gaussian mixture model refers to the set of all color slice regions. • An ear can be a mixture of Gaussian models and mixture model deals with color features in the color feature space. • For segmentation of color features in the feature space in terms of pixels in detected ear image based on the probabilities of identical color spaces, vector quantization is applied to cluster the color features of pixels. • Vector quantization [15] can be considered as a fitting model where the clusters are represented by conditional density functions.
  • 10. May 20 - 24, 2010 | Busan, Korea 10 Ear Modeling using GMM: • In this fitting model, predetermined set of probabilities are the weights. • Data contained within vector quantization framework can be fitted with Gaussian mixture model and the probability density function of a dataset is represented as a collection of Gaussians. This convention can be represented by the following equation: )|()( 1 ixfPxf N i i∑= = where N is the number of clusters or slice regions in ear image, Pi is the prior probability of cluster i and f(x/i) is the probability density function of cluster i.
  • 11. May 20 - 24, 2010 | Busan, Korea 11 Contd… The conditional probability density function f(x/i) can be represented as 2 1 2 1 ||)2( ))()( 2 1 exp( )|( ∑ ∑ − −−− = i P i i t i mxmx ixf π where x ε RP, mi and ∑i are the mean and covariance matrix of cluster i respectively. To determine the maximum likelihood parameters of a mixture of i Gaussians, the Expectation- Maximization (EM) algorithm [18] is used while the Minimum Description Length (MDL) principle is used to select the values of i ranging from 3 to 6.
  • 12. May 20 - 24, 2010 | Busan, Korea 12 K-L Divergence for Color Similarity Measurement: • Kullback-Leibler (K-L) divergence [14] is given as non-symmetric distance measure between probability distributions. • In computer vision and pattern classification, it is often needed to compute the similarity between two images or coherent regions of two images. It is performed by matching the spatial features or color features of the images. • K-L divergence measures the theoretic criterion that gives a dissimilarity score between the probabilities densities of two images or regions of images. • It measures the expected number of extra bits required to code samples from one probability distribution when using a code based on another probability distribution, rather than using a code based on the first distribution.
  • 13. May 20 - 24, 2010 | Busan, Korea 13 Contd… • Therefore, the first distribution model represents the "true" distribution of data, observations or a precise calculated theoretical distribution. The second probability distribution measure typically represents a theory, model, description, or approximation of the first one. • Once Gaussian mixture models [13] for color pixels have been formed in the cropped ear images, K-L divergence is used for keep color consistency in the coherent color slice regions independently and is also used for finding similarity among the ear images in terms of mixture of Gaussian models. • The K-L divergence can be defined between two probability density functions p(x) and q(x) found from two color ear images, ∑= x def xq xp xpqpKL )( )( log)()||(
  • 14. May 20 - 24, 2010 | Busan, Korea 14 SIFT Features Extraction: • The Scale Invariant Feature Transform (SIFT) descriptor [8] is invariant to image rotation, scaling, partly illumination changes and the 3D camera view. – The SIFT descriptor detects feature points efficiently through a staged filtering approach that identifies stable points in the scale- space of the resulting image pyramid. – Local feature points are extracted through selecting the candidates for feature points by searching peaks in the scale-space from a DoG function. – Further the feature points are localized using the measurement of their stability and assign orientations based on local image properties. – Finally, the feature descriptors, which represent local shape distortions and illumination changes, are determined.
  • 15. May 20 - 24, 2010 | Busan, Korea 15 Contd… – Prior to feature extraction from color slice regions, slice regions are converted into grayscale slice regions by using the technique presented in [16]. – The ear model is normalized by histogram equalization and then SIFT features [8] are extracted from the color slice regions. – Each feature point contains four types of information – spatial location (x, y), scale (S), orientation (θ) and Keypoint descriptor (K). – For the experiment, only keypoint descriptor [8] information has been used which consists of a vector of 128 elements representing neighborhood intensity changes of each keypoint.
  • 16. May 20 - 24, 2010 | Busan, Korea 16 Fusion Strategy of Keypoint Features: • In the proposed ear recognition model, detected SIFT features from color-segmented slice regions are fused together by concatenation and Dempster-Shafer decision theory. • The keypoints are extracted from different slice regions are taken to make an augmented group of features for both the reference ear model and the probe ear model. • The proposed fusion strategies use feature level fusion approaches which are used to fuse the feature sets obtained from different color segmented slice regions. • Fusion using concatenation approach: – In order to obtain fused sets of features for both the reference and the probe models, the keypoints are detected in varying number for each segment region as K1, K2, K3,… KS. – Now, an augmented set is obtained DS of SIFT features by concatenation as follows }.....{ 321 SKKKKDS ∪∪∪∪=
  • 17. May 20 - 24, 2010 | Busan, Korea 17 Contd… – The feature set DS represents the proximity among detected SIFT features of the color slice regions. – Finally, the final matching distance Dfinal (DSprobe, DSreference) is computed on the basis of the number of keypoints paired between two sets of features and is given by Ψ≤−= ∑∈∈ referenceprobe DjDi jreferenceiprobefinal KDSKDSD , 2 ))()(( where DSprobe and DSreference are the concatenated feature sets for both the probe model and the reference model and ѱ is the threshold determined from a subset of database. As for the matching threshold, this ear set is disjoint from the image sets used for testing and validation.
  • 18. May 20 - 24, 2010 | Busan, Korea 18 Contd… • Fusion using Dempster-Shafer Decision Theory: – The Dempster-Shafer decision theory [3] is used to integrate the detected keypoint features obtained from individual slice regions. – It is based on combining the evidences obtained from different sources to compute the probability of an event. – This is obtained by combining three elements: the basic probability assignment function (bpa), the belief function (bf) and the plausibility function (pf). Details of Dempster-Shafer theory can be found in [3]. – Let ГSR1, ГSR2, ГSR3,….,ГSRn be the feature sets obtained from the n numbers of slice regions. – The dimensions of feature points in all sets of slice regions may be different. – To make the features sets of equal dimensions, the features set with less numbers of features points added with zero to the remaining points and make it equal length with other feature sets.
  • 19. May 20 - 24, 2010 | Busan, Korea 19 Contd… – Now, in order to obtain the transformed combine feature set Dempster combination rule is applied. – Also, let m(ГSR1), m(ГSR2), m(ГSR3),….,m(ГSRn) be the bpa functions for the Belief measures Bel(ГSR1), Bel(ГSR2), Bel(ГSR3),…, Bel(ГSRn) for the n numbers of slice regions respectively. – Then the Belief probability assignments (bpa) can be combined together to obtained a Belief committed to a feature set according to the following combination rule or orthogonal sum rule ., )()(1 )()( )()()( 121 21 21 1 21 21 ∅≠ ΓΓ− ΓΓ =Γ⊕Γ= ∑ ∑ ∅≠Γ∩Γ =Γ∩Γ C mm mm mmCm SRSR SRSR SRSR C SRSR SRSR ., )()(1 )()( )()()( 243 43 43 2 43 2 43 ∅≠ ΓΓ− ΓΓ =Γ⊕Γ= ∑ ∑ ∅≠Γ∩Γ =Γ∩Γ C mm mm mmCm SRSR SRSR SRSR C SRSR SRSR
  • 20. May 20 - 24, 2010 | Busan, Korea 20 Contd… • Let, m(C1), m(C2),….., m(Cn) be obtained from different sets of pairs of features. They can be further fused by the following equation ., )()(1 )()( )()()( )1( )1( )1( )1( )1( ∅≠ ΓΓ− ΓΓ =Γ⊕Γ= ∑ ∑ ∅≠Γ∩Γ − =Γ∩Γ − − − − nSRnnSR C SRnnSR SRnnSR n C mm mm mmCm SRnnSR n SRnnSR )(....)()()( 321 nCmCmCmCmS ⊕⊕⊕⊕= • The denominators of Equations from (9) to (11) denote the normalizing factors which denote the art of belief assignments. • The notation denotes the Dempster combination rule. The combined value S represents the transformed feature value and representative feature of all keypoint features.
  • 21. May 20 - 24, 2010 | Busan, Korea 21 Contd… • Let, S1 and S2 be the final transformed feature values obtained from reference ear model and probe ear model respectively. • Finally, the final matching distance D’final’ (S1, S2) is computed on the basis of the number of keypoints paired between two sets of features. • The similarity score can be computed as follows ∑ −= 2 12' )(' SSD final The final decision of user acceptance and rejection can be established by applying the threshold Ф to the final matching score D’final’.       Φ≤ = otherwisereject Difaccept decision final , ', ' where Ф is a predefined threshold value.
  • 22. May 20 - 24, 2010 | Busan, Korea 22 Experimental Results: • The proposed identification technique has been tested on IIT Kanpur ear database [12]. • The database consists of 800 ear images of 400 individuals and all the frontal view ear images are considered for evaluation. • The ear images are taken under controlled environment in different sessions. The ear viewpoints are consistently kept neutral and the ear images are downscaled to 200×240 pixels with 500 dpi resolution. • High resolution ear image helps to increase the SIFT keypoints during feature extraction. Ear images are acquired by high resolution digital camera. • The entire database is divided into two different groups such as reference and probe. For reference model, single ear image is enrolled for each individual. • Therefore, 400 ear images are considered for training session. and the remaining images are considered for testing and evaluation.
  • 23. May 20 - 24, 2010 | Busan, Korea 23 Contd… Table 1. Identification Rates for the Proposed Ear Identification Systems 98.25After color segmentation (DS theory based fusion rule) 94.75After color segmentation (Concatenation fusion rule) 92.5Prior to color segmentation IDENTIFICATION RATE (%) METHODS ↓ 0 10 20 30 40 50 60 70 80 90 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99 1 <--- Rank ---> <---IdentificationProbability---> Cumulative characteristics (CMC) Curves DS Theory Based Fusion(After Color Segmentation) Concatenation Fusion Rule(After Color Segmentation) Prior to Color Segmentation Figure 2. Cumulative Match Characteristics Curves (CMC) of the Proposed Methods
  • 24. May 20 - 24, 2010 | Busan, Korea 24 Conclusion: • This paper has proposed an efficient ear identification and verification system which uses SIFT descriptor for feature extraction from color similarity slice regions. • The experiments have been conducted in two different sessions. In the first session, results are obtained prior to segmentation of ear images and are supported through the two distance metrics. In the next session, segmented regions are considered for feature extraction and matching SIFT keypoint features. • It has used two different fusion rules, viz. concatenation and Dempster combination rule for fusing the keypoint features extracted from slice regions. • During identification in the second session, the system achieves with an accuracy of more than 97%. • This proves that the system can be deployed to high security applications where single modality can be used with efficient and cost effective manner.
  • 25. May 20 - 24, 2010 | Busan, Korea 25 Bibliography: • J.D. Bustard, and M.S. Nixon, “Robust 2D ear registration and recognition based on SIFT point matching,” Proc. of the International Conference on Biometrics: Theory, Applications, and Systems, pp. 1-6, 2008. • Y. Liu, Z. Mu, and L. Yuan, “Application of kernel function based fisher discriminant analysis algorithm in ear recognition,” Measurements and Control, vol. 22, no. 8, pp. 304 – 306, 2006. • D.R. Kisku, M. Tistarelli, J.K. Sing, J. K. and P. Gupta, "Face recognition by fusion of local and global matching scores using DS theory: An evaluation with uni-classifier and multi-classifier paradigm,” Proc. IEEE Computer Vision and Pattern Recognition (CVPR) Workshop on Biometrics, pp. 60-65, 2009. • M. Burge, and W. Burger, “Ear biometrics,” BIOMETRICS: Personal Identification in a Networked Society, pp.273 – 286, 1999. • D.J. Hurley, M.S. Nixon, and J.N. Carter, “A new force field transform for ear and face recognition,” Proc. of the IEEE International Conference on Image Processing, pp. 25 – 28, 2000. • K. Chang, K. Bowyer, S. Sarkar, and B. Victor, “Comparison and combination of ear and face images in appearance-based biometrics,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 25, no. 9, pp. 1160 – 1165, 2003. • T. Lindeberg, "Feature detection with automatic scale selection," International Journal of Computer Vision, vol. 30, no. 2, pp. 79 – 116, 1998. • D.G. Lowe, "Object recognition from local scale-invariant features," Proc. International Conference on Computer Vision, pp. 1150 – 1157, 1999. • H. Chen, and B. Bhanu, “Human ear recognition in 3D,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 29, pp. 718 – 737, 2007. • Y. Wang, Z. Mu, and H. Zeng, “Block-based and multi-resolution methods for ear recognition using wavelet transform and uniform local binary patterns,” Proc. International Conference on Pattern Recognition, pp. 1-4, 2008. • G.S. Badrinath, and P. Gupta, “Feature level-fused ear biometric system,” Proc. International Conference on Advances in Pattern Recognition, pp. 197 – 200, 2009.
  • 26. May 20 - 24, 2010 | Busan, Korea 26 Bibliography: Contd… • D.R. Kisku, P. Gupta, and J.K. Sing, “Feature level fusion of biometric cues: Human identification with Doddington's caricature,” Proc. International Conference on Security Technology, Communications in Computer and Information Sciences, Springer-Verlag, pp. 157-164, 2009. • R. M. Gray, and T. Linder, “Mismatch in high rate entropy constrained vector quantization,” IEEE Transactions on Information Theory, vol. 49, pp. 1204 – 1217, 2003. • J. Goldberger, S. Gordon, and H. Greenspan, “An efficient image similarity measure based on approximations of K-L Divergence between two Gaussian mixtures,” Proc. International Conference on Computer Vision, pp. 1-5, 2003. • A. Gersho, and R. Gray, “Vector Quantization and Signal Compression,” Kluwer Academic Press, 1992. • M. Grundland, and N. A. Dodgison, “Decolorize: Fast, contrast enhancing, color to gray scale conversion,” Pattern Recognition, vol. 40, no. 11, pp.2891-2896, 2007.
  • 27. May 20 - 24, 2010 | Busan, Korea 27 Thank You ! Questions ??? Contact person drkisku@ieee.org