WWW.DISSERTATION.XLIBX.INFO
FREE ELECTRONIC LIBRARY - Dissertations, online materials
 
<< HOME
CONTACTS



Pages:     | 1 |   ...   | 7 | 8 || 10 | 11 |   ...   | 33 |

«Ear Recognition Biometric Identification using 2- and 3-Dimensional Images of Human Ears Anika Pflug Thesis submitted to Gjøvik University College ...»

-- [ Page 9 ] --

Abate et al. [1] use a generic Fourier descriptor for rotation and scale invariant feature representation. The image is transformed into a polar coordinate system and then transformed into frequency space. In order to make sure, that the centroid of the polar coordinate system is always at the same position, the ear images have to be aligned before they can be transformed into the polar coordinate system. The concha serves as a reference point for the alignment step, such that the center point of the polar coordinate system is always located in the concha region. The approach was tested on a proprietary dataset, which contains 282 ear images in total. The images were taken on two different days and in different roll and yaw poses. The rank-1 performance of the Fourier descriptor varies depending on the pose angle. For 0 degrees pose variation the rank-1 performance is 96%, but if different poses are included in the experiments, it drops to 44% for 15 degrees and 19% for 30 degres.

In the work of Fooprateepsiri and Kurutach exploit the concepts of multi-resolution Trace transform and Fourier transform. The input images from the CMU PIE database are serialized by using the trace transform and stored in a feature vector. The advantage of the trace transform is that the resulting feature vector is invariant to rotations and scale.

Furthermore Fooprateepsiri and Kurutach show that their descriptor is also robust against

3. EAR BIOMETRICS: A SURVEY OF DETECTION, FEATURE EXTRACTION AND

RECOGNITION METHODS

pose variations. In total they report a rank-1 performance of 97%.

Sana et al. use selected wavelet coefficients extracted during Haar-Wavelet compression for feature representation [169]. While applying the four level wavelet transform several times on the ear image, for each iteration they store one of the derived coefficients in a feature vector. The reported accuracy of their algorithm is 96% and was achieved on the basis of the IITK database and on the Saugor database (350 subjects).

A feature extraction system called PIFS is proposed by De Marisco et al. [60]. PIFS measures the self-similarity in an image by calculating affine translations between similar sub regions of an image. In order to make their system robust to occlusion, De Marisco et al.

divided the ear image into equally large tiles. If one tile is occluded, the other tiles still contain a sufficiently distinctive set of features. De Marisco et al. could show that their approach is superior to other feature extraction methods under the presence of occlusion. The experiments of De Marisco et al. have been conducted in order to assess the system performance in different occlusion scenarios. The basis for these tests was the UND collection E and the first 100 subjects of the FERET database. If occlusion occurs on the reference image, a rank-1 performance of 61% (compared to 40% on average with other feature extraction methods) is reported. Without occlusion, the rank-1 performance is 93%.

Moment invariants are a statistical measure for describing specific properties of a shape.

Wang et al. [191] compose six different feature vectors by using seven moment invariants. They also show that each of the moment invariants is robust against changes in scale and rotation. The feature vectors are used as the input for a back propagation neural network which is trained to classify the moment invariant feature sets. Based on a proprietary database of 60 ear images, they report a rank-1 performance of 91.8%. In [189] Wang and Yuan compare the distinctiveness of different feature extraction methods on the USTB I database. They compare the rank-1 performance of Fourier descriptors, Gabor-Transform, Moment Invariants and statistical features and come to the conclusion that the highest recognition rate can be achieved by using moment invariants and Gabor transform. For both feature extraction methods Wang and Yuan report a rank-1 performance or 100%.

3.4.2 Local Descriptors Scale invariant Feature Transform (SIFT) is known to be a robust way for landmark extraction even in images with small pose variations and varying brightness conditions [118].

SIFT landmarks contain a measure for local orientation; they can also be used for estimating the rotation and translation between two normalized ear images. Bustard et al. showed that SIFT can handle pose variations up to 20 degrees [39]. However it is not a trivial to assign a SIFT landmark with its exact counterpart, especially in the presence of pose variations. In highly structured image regions, the density and redundancy of SIFT landmarks is so high, that exact assignment is not possible. Hence the landmarks have to be filtered before the actual comparison can start. Arbab-Zavar et al. [19] as well as Badrinath and Gupta [21] therefore train a reference landmark model, which only contains a small number of non-redundant landmarks. This landmark model is used for filtering the SIFT landmarks, which were initially detected in the probe and reference ear. Having the filtered landmarks it is possible to assign each of the landmarks with its matching counterpart. Figure 3.7 shows an example for SIFT landmarks extracted from ear images, which were used as training data for the reference landmark model in the work of Arbab-Zavar et al.. Because Arbab-Zavar et al. also used the XM2VTS database for evaluation, their results can be directly compared to the rank-1 performance reported by Bustard and Nixon.





Arbab-Zavar et al. achieved a rank-1 performance of 91.5%. With the more recent approach by Bustard and Nixon the performance could be improved to 96%. Using the IIT Delhi database Kumar et al. report a GAR of 95% and a FAR of 0.1% when using SIFT feature points.

Kisku et al. address the problem of correct landmark assignment by decomposing the ear image into different color segments [102]. SIFT landmarks are extracted from each segD EAR RECOGNITION ment separately, which reduces the chance of assigning SIFT landmarks that are not representing the same features. Using this approach, Kisku et al. achieve a rank-1 performance of 96.93%.

A recent approach by Prakash and Gupta [156] fuses Speeded Up Robust Features (SURF) [23] feature points from different images of the same subject. They propose to use several input images for enrolment and to store all SURF feature points in the fused feature vector, which could be found in the input images. These feature sets are then used for training a nearest neighbor classifier for assigning two correlated feature points. If the distance between two SURF feature points is less than a trained threshold, they are considered to be correlated. The evaluation of this approach was carried out on the UND collection E and the two subsets of the IIT Kanpur database. Prakash and Gupta tested the influence of different parameters for SURF features and for the nearest neighbor classifier. Depending on the composition of the parameters the EER varies between 6.72% and 2.25%.

Choras proposes a set of geometric feature extraction methods inspired by the work of Iannarelli [52]. He proposes four different ways of feature location in edge images. The concentric circles method uses the concha as reference points for a number of concentric circles with predefined radii. The intersection points of the circles and the ear contours are used as feature points (see Figure 3.7.). An extension of this is the contour tracing method, which uses bifurcations, endpoints and intersecting points between the ear contours as additional features. In the angle representation approach, Choras draws concentric circles around each center point of an edge and uses the angles between the center point and the concentric circles intersecting points for feature representation. Finally the triangle ratio method determines the normalized distances between reference points and uses them for ear description. Choras conducted studies on different databases where he reported recognition rates between 86.2% and 100% on a small database off 12 subjects and a false reject rate between 0% and 9.6% on a larger database with 102 ear images.

Similar approaches which are using the aspect ratio between reference points on the ear contours are proposed by Mu et al. with a rank-1 performance of 85% on the USTB II database [129] and Rahman et al. [163]. Rahman et al. evaluated their approach on a database, which consists of 350 images from 100 subjects. They report a rank-1 performance of 90%. For images, which were taken on different days the rank-1 performance dropped to 88%.

Local binary patterns (LBP) are a technique for feature extraction on the pixel level.

LBP encode the local neighborhood of a pixel by storing the difference between the examined pixel and its neighbors. Guo et al. extract LBP from the raw ear images and create histograms describing the distribution of the local LBP. Then a cellular neural network is trained to distinguish between the LBP of different subjects in the USTB II database [72].

In the by Wang and Yan [193] the dimensionality of the feature vector is reduced with linear discriminant analysis before a Euclidean distance measure quantifies the similarity of two feature vectors. Wang and Yan evaluated their approach on the USTB II dataset and report a rank-1 performance of 100%.

3.4.3 Hybrid Approaches The approach of Jedges and Mate is twofold [97]. In a first feature extraction step they generate an average edge model from a set of training images. These edges represent the outer helix contour as well as the contours of the antihelix, the fossa triangularis and the concha. Subsequently each image is enrolled by deforming the ear model until it fits the actual edges displayed in the probe ear image. The deformation parameters, which were necessary for the transformation, are the first part of the feature vector. The feature vector is completed by adding additional feature points lying on intersections between a predened set of axes and the transformed main edges. The axes describe the unique outline of ear. Figure 3.7 shows the edge enhanced images with fitted contours together with the

3. EAR BIOMETRICS: A SURVEY OF DETECTION, FEATURE EXTRACTION AND

RECOGNITION METHODS

additional axes for reference point extraction. They report an EER of 5.6% using a database with cropped images and without pose variations.

Liu et al. combine front and backside view of the ear by extracting features using the triangle ratio method and Tchebichef moment descriptors [115]. Tchebichef moments are a set of orthogonal moment functions based on discrete Tchebichef polynomials and have been introduced as a method for feature representation in 2001 [131]. The backside of the ear is described by a number of lines that are perpendicular to the longest axis in the ear contour. These lines measure the local diameter of the auricle at predefined points. The rank-1 performance of this combined approach is reported to be 97.5%. If only the front view is used, the rank-1 performance is 95% and for the backside images, Liu et al. report 86.3% rank-1 performance.

Lu et al. [119] as well as Yuan and Mu [207] use the active shape model for extracting the outline of the ear. Lu et al. are using manually cropped ear images from 56 subjects in different poses. A feature extractor stores selected points on the outline of the ear together with their distance to the tragus. Before applying a linear classifier, the dimensionality of the feature vectors is reduced by principal component analysis (PCA). Lu et al. compare the rank-1 performance of pipelines where only the left or the right ear was used for identification and also show that using both ears increases the rank-1 performance from 93.3% to 95.1%. In the USTB III database Yuan and Mu report a rank-1 performance of 90% if the head rotation is lower than 15 degrees. For rotation angles between 20 degrees and 60 degrees the rank-1 performance drops to 80%.

3.4.4 Classifiers and Statistical Approaches Victor et al. were the first research group to transfer the idea of using the Eigen space from face recognition to ear recognition [180]. They reported that the performance of the ear as a feature is inferior to the face. This may be due to the fact that in their experiments Victor et al. considered the left and the right ear to be symmetric. They used the one ear for training and the other ear for testing, which could have lowered the performance of PCA in this case. The reported rank-1 performance is 40%. With a rank-1 performance of 72.2% in the UND collection E, Chang et al. [46] report a significantly better performance than Victor et al.. Alaraj et al. [12] published another study, where PCA is used for feature representation in ear recognition. In their approach a multilayer feed forward neural network was trained for classification of the PCA based feature components. The observed a rank-1 performance of 96%, and hence improved the previous results by Victor et al. and Chang et al.. However it should be noticed that this result is only based on a subset of one of the UND collections, which consists of 85 ear images from 17 subjects.

Zhang and Mu conducted studies on the effectiveness of statistical methods in combination with classifiers. In [215] they show that independent component analysis (ICA) is more effective on the USTB I database than PCA. They first used PCA and ICA for reducing the dimensionality of the input images and then trained an SVM for classifying the extracted feature vectors. Furthermore the influence of different training set sizes on the performance was measured. Depending on the size of the training set the rank-1 performance for PCA varies between 85% and 94.12%, whereas the rank-1 performance for ICA varies between 91.67% and 100%.

Xie and Mu [198] propose an improved locally linear embedding (LLE) algorithm for reducing the dimensionality of ear features. LLE is a technique for projecting high-dimensional data points into a lower dimensional coordinate system while preserving the relationship between the single data points. This requires the data points to be labeled in some way, so that their relationship is fixed. The improved version of LLE by Xie and Mu eliminated the problem by using a different distance function. Further Xie and Mu show, that LLE is superior to PCA and Kernel PCA, if the input data contains pose variations. Their studies were conducted on the USTB III database showed that the rank-1 performance of regular



Pages:     | 1 |   ...   | 7 | 8 || 10 | 11 |   ...   | 33 |


Similar works:

«Welcome Home to Walden Pond The purpose of this handbook is to let you know about our community. For over three decades, HHHunt has been a recognized leader in real estate development throughout Maryland, Virginia, North Carolina, Tennessee, and South Carolina. Our founder, Harry H. Hunt III, attributes the company’s strong and consistent growth to its core management philosophy. “We value all of our employees and respect their ideas,” he says. “They know that the best idea, no matter...»

«http://unllib.unl.edu/LPP/ Library Philosophy and Practice 2011 ISSN 1522-0222 Doctoral Research in Library and Information Science by Pakistani Professionals: An Analysis Dr. Rais Ahmed Samdani Associate Professor Department. of Library & Information Sciences University of Sargodha, Pakistan Dr. Rubina Bhatti Assistant Professor Department. of Library & Information Sciences The Islamia University of Bahawalpur, Pakistan Introduction Research is a process of enquiry that draws data from the...»

«BEAUTY PAGEANTS IN NEOLIBERAL CHINA: A FEMINIST MEDIA STUDY OF FEMININE BEAUTY AND CHINESE CULTURE By MENG ZHANG A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA © 2013 Meng Zhang To my Parents ACKNOWLEDGMENTS I would like to give my deepest gratitude to my advisor and committee chair, Dr. Cory Armstrong, who has shown me the dedication, spirit, and substance of a...»

«Self-directed learning: managing yourself and your working relationships TIME MANAGEMENT Many years ago, before I became involved in management development, I attended a twoday time management course. It was a well run programme, and I found it useful. For me, the mark of a successful course is that you as a participant gain one thing that you are still using years later. I’ll tell you shortly what I took from that course. However, even as the course was unfolding I realised that we were...»

«i CUSTOMER SATISFACTION AND LOYALTY IN THE APPLICATION OF THE ALL-INCLUSIVE HOLIDAY CONCEPT AT THE KENYAN COAST BY MARY. M. MUTISYA (B.ED; M.SC) REG NO. H87/10857/2007 A THESIS SUBMITTED IN FULFILMENT OF THE REQUIREMENTS FOR THE AWARD OF THE DEGREE OF DOCTOR OF PHILOSOPHY (HOSPITALITY MANAGEMENT) IN THE SCHOOL OF HOSPITALITY AND TOURISM MANAGEMENT OF KENYATTA UNIVERSITY NOVEMBER 2011 ii DECLARATION This Thesis is my original work and has not been presented for a degree in any other university...»

«EXPERIENTIAL LEARNING IN UNDERGRADUATE PHARMACY CURRICULUM: A CASE STUDY OF COOPERATIVE EXPERIENCE OF PHARMACY STUDENTS by Kum Tong Certina Ho A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy Department of Curriculum, Teaching and Learning Ontario Institute for Studies in Education University of Toronto © Copyright by Kum Tong Certina Ho (2016) EXPERIENTIAL LEARNING IN UNDERGRADUATE PHARMACY CURRICULUM: A CASE STUDY OF CO-OPERATIVE EXPERIENCE OF...»

«A MODEL BASED APPROACH TO THE ANALYSIS OF INTERSECTION CONFLICTS AND COLLISION AVOIDANCE SYSTEMS by Kazutoshi Nobukawa A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy (Mechanical Engineering) in The University of Michigan 2011 Doctoral Committee: Professor Timothy J. Gordon, Chair Professor Jessy W. Grizzle Professor Huei Peng Professor Noel C. Perkins © Kazutoshi Nobukawa 2011 DEDICATION To my parents. ii ACKNOWLEDGMENTS This...»

«THE ROLE OF IL-21 AND IL-17 IN REGULATING FOLLICULAR T HELPER CELLS IN GERMINAL CENTER RESPONSE OF AUTOIMMUNITY by YANNA DING DR. ZDENEK HEL, COMMITTEE CHAIR DR. PETER D. BURROWS DR. JOHN D. MOUNTZ, MENTOR DR. MOON NAHM DR. CHANDER RAMAN A DISSERTATION Submitted to the graduate faculty of The University of Alabama at Birmingham, in partial fulfillment of the requirements for the degree of Doctor of Philosophy BIRMINGHAM, ALABAMA 2013 THE ROLE OF IL-21 AND IL-17 IN REGULATING FOLLICULAR T HELPER...»

«DIPLOMA THESIS Linking Climate Change with Food Security in the highlands of Khyber Pakhtunkhwa, Northwest Pakistan Presented by: Martin Kienzler Julius-Maximilians-Universität Würzburg Faculty of Philosophy I Institute of Geography and Geology Physical Geography DIPLOMA THESIS Linking Climate Change with Food Security in the highlands of Khyber Pakhtunkhwa, Northwest Pakistan Presented by Martin Kienzler, born on June, 25th 1985 in Titisee-Neustadt to obtain the degree of DIPLOMGEOGRAPH...»

«An Ambivalent, Postphenomenological Philosophy of Technology What Things Do: Reflections on Technology, Agency, and Design. By Peter-Paul Verbeek. Translated by Robert P. Crease. University Park: Pennsylvania State University Press, 2005. 249 pp. ISBN: 0271025409. Paperback. $30 Review by Robert Rosenberger With What Things Do: Reflections on Technology, Agency, and Design, Peter-Paul Verbeek (2005) provides an in-depth review and synthesis of the history of the field of philosophy of...»

«THE EFFECT OF COMBUSTION CHAMBER DEPOSITS ON HEAT TRANSFER AND COMBUSTION IN A HOMOGENEOUS CHARGE COMPRESSION IGNITION ENGINE by Orgun A. Güralp A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy (Mechanical Engineering) in The University of Michigan 2008 Doctoral Committee: Professor Dionissios N. Assanis, Co-Chair Associate Research Professor Zoran S. Filipi, Co-Chair Professor Volker Sick Professor James F. Driscoll Tang-Wei Kuo,...»

«On Ethical Thoughtfulness Item type text; Electronic Dissertation Authors Matteson, Jason Kent Publisher The University of Arizona. Rights Copyright © is held by the author. Digital access to this material is made possible by the University Libraries, University of Arizona. Further transmission, reproduction or presentation (such as public display or performance) of protected items is prohibited except with permission of the author. Downloaded 23-Nov-2016 06:46:30 Link to item...»





 
<<  HOME   |    CONTACTS
2016 www.dissertation.xlibx.info - Dissertations, online materials

Materials of this site are available for review, all rights belong to their respective owners.
If you do not agree with the fact that your material is placed on this site, please, email us, we will within 1-2 business days delete him.