1 Fingerprint Classification sections 5.3 - 5.5 Fingerprint matching using transformation parameter clustering R. Germain et al, IEEE And Fingerprint Identification.

Slides:



Advertisements
Similar presentations
Discrimination and Classification. Discrimination Situation: We have two or more populations  1,  2, etc (possibly p-variate normal). The populations.
Advertisements

Alignment Visual Recognition “Straighten your paths” Isaiah.
Order Structure, Correspondence, and Shape Based Categories Presented by Piotr Dollar October 24, 2002 Stefan Carlsson.
Road-Sign Detection and Recognition Based on Support Vector Machines Saturnino, Sergio et al. Yunjia Man ECG 782 Dr. Brendan.
Principal Component Analysis Based on L1-Norm Maximization Nojun Kwak IEEE Transactions on Pattern Analysis and Machine Intelligence, 2008.
Presented by Xinyu Chang
Extended Gaussian Images
3D Shape Histograms for Similarity Search and Classification in Spatial Databases. Mihael Ankerst,Gabi Kastenmuller, Hans-Peter-Kriegel,Thomas Seidl Univ.
Fingerprint Minutiae Matching Algorithm using Distance Histogram of Neighborhood Presented By: Neeraj Sharma M.S. student, Dongseo University, Pusan South.
A 3-D reference frame can be uniquely defined by the ordered vertices of a non- degenerate triangle p1p1 p2p2 p3p3.
An Approach to Evaluate Data Trustworthiness Based on Data Provenance Department of Computer Science Purdue University.
The Statistics of Fingerprints A Matching Algorithm to be used in an Investigation into the Reliability of the Use of Fingerprints for Identification Bob.
66: Priyanka J. Sawant 67: Ayesha A. Upadhyay 75: Sumeet Sukthankar.
A Versatile Depalletizer of Boxes Based on Range Imagery Dimitrios Katsoulas*, Lothar Bergen*, Lambis Tassakos** *University of Freiburg **Inos Automation-software.
Multiple Criteria for Evaluating Land Cover Classification Algorithms Summary of a paper by R.S. DeFries and Jonathan Cheung-Wai Chan April, 2000 Remote.
Content Based Image Clustering and Image Retrieval Using Multiple Instance Learning Using Multiple Instance Learning Xin Chen Advisor: Chengcui Zhang Department.
Robust and large-scale alignment Image from
Instructor: Dr. G. Bebis Reza Amayeh Fall 2005
A Study of Approaches for Object Recognition
Symmetric hash functions for fingerprint minutiae
Evaluating Hypotheses
Pores and Ridges: High- Resolution Fingerprint Matching Using Level 3 Features Anil K. Jain Yi Chen Meltem Demirkus.
Fitting a Model to Data Reading: 15.1,
Object Recognition Using Geometric Hashing
A unified statistical framework for sequence comparison and structure comparison Michael Levitt Mark Gerstein.
Comparison and Combination of Ear and Face Images in Appearance-Based Biometrics IEEE Trans on PAMI, VOL. 25, NO.9, 2003 Kyong Chang, Kevin W. Bowyer,
Automatic Image Alignment (feature-based) : Computational Photography Alexei Efros, CMU, Fall 2006 with a lot of slides stolen from Steve Seitz and.
Recognition of object by finding correspondences between features of a model and an image. Alignment repeatedly hypothesize correspondences between minimal.
Geometric Hashing Visual Recognition Lecture 9 “Answer me speedily” Psalm, 17.
Ashish Uthama EOS 513 Term Paper Presentation Ashish Uthama Biomedical Signal and Image Computing Lab Department of Electrical.
Face Recognition Using Neural Networks Presented By: Hadis Mohseni Leila Taghavi Atefeh Mirsafian.
Principal Component Analysis (PCA) for Clustering Gene Expression Data K. Y. Yeung and W. L. Ruzzo.
October 8, 2013Computer Vision Lecture 11: The Hough Transform 1 Fitting Curve Models to Edges Most contours can be well described by combining several.
CSE 185 Introduction to Computer Vision
Distinctive Image Features from Scale-Invariant Keypoints By David G. Lowe, University of British Columbia Presented by: Tim Havinga, Joël van Neerbos.
AdvisorStudent Dr. Jia Li Shaojun Liu Dept. of Computer Science and Engineering, Oakland University 3D Shape Classification Using Conformal Mapping In.
Active Learning for Class Imbalance Problem
CS 376b Introduction to Computer Vision 04 / 29 / 2008 Instructor: Michael Eckmann.
Presented by Tienwei Tsai July, 2005
Shape Matching for Model Alignment 3D Scan Matching and Registration, Part I ICCV 2005 Short Course Michael Kazhdan Johns Hopkins University.
MESA LAB Multi-view image stitching Guimei Zhang MESA LAB MESA (Mechatronics, Embedded Systems and Automation) LAB School of Engineering, University of.
Intelligent Vision Systems ENT 496 Object Shape Identification and Representation Hema C.R. Lecture 7.
Generalized Hough Transform
Features-based Object Recognition P. Moreels, P. Perona California Institute of Technology.
Mingyang Zhu, Huaijiang Sun, Zhigang Deng Quaternion Space Sparse Decomposition for Motion Compression and Retrieval SCA 2012.
2005/12/021 Content-Based Image Retrieval Using Grey Relational Analysis Dept. of Computer Engineering Tatung University Presenter: Tienwei Tsai ( 蔡殿偉.
Event retrieval in large video collections with circulant temporal encoding CVPR 2013 Oral.
2005/12/021 Fast Image Retrieval Using Low Frequency DCT Coefficients Dept. of Computer Engineering Tatung University Presenter: Yo-Ping Huang ( 黃有評 )
Geometric Hashing: A General and Efficient Model-Based Recognition Scheme Yehezkel Lamdan and Haim J. Wolfson ICCV 1988 Presented by Budi Purnomo Nov 23rd.
CS 376b Introduction to Computer Vision 04 / 28 / 2008 Instructor: Michael Eckmann.
CVPR2013 Poster Detecting and Naming Actors in Movies using Generative Appearance Models.
EE 7740 Fingerprint Recognition. Bahadir K. Gunturk2 Biometrics Biometric recognition refers to the use of distinctive characteristics (biometric identifiers)
An Approximate Nearest Neighbor Retrieval Scheme for Computationally Intensive Distance Measures Pratyush Bhatt MS by Research(CVIT)
1 An Efficient Classification Approach Based on Grid Code Transformation and Mask-Matching Method Presenter: Yo-Ping Huang.
Chapter 13 (Prototype Methods and Nearest-Neighbors )
Iterative similarity based adaptation technique for Cross Domain text classification Under: Prof. Amitabha Mukherjee By: Narendra Roy Roll no: Group:
October 16, 2014Computer Vision Lecture 12: Image Segmentation II 1 Hough Transform The Hough transform is a very general technique for feature detection.
Feature Selction for SVMs J. Weston et al., NIPS 2000 오장민 (2000/01/04) Second reference : Mark A. Holl, Correlation-based Feature Selection for Machine.
776 Computer Vision Jan-Michael Frahm Spring 2012.
FINGER PRINT RECOGNITION USING MINUTIAE EXTRACTION FOR BANK LOCKER SECURITY Presented by J.VENKATA SUMAN ECE DEPARTMENT GMRIT, RAJAM.
Finding Near-Duplicate Web Pages: A Large-Scale Evaluation of Algorithms By Monika Henzinger Presented.
Improving the Performance of Fingerprint Classification
Principal Component Analysis (PCA)
Discrimination and Classification
Fitting Curve Models to Edges
K Nearest Neighbor Classification
Hu Li Moments for Low Resolution Thermal Face Recognition
Application: Geometric Hashing
Geometric Hashing: An Overview
Handwritten Characters Recognition Based on an HMM Model
Presentation transcript:

1 Fingerprint Classification sections Fingerprint matching using transformation parameter clustering R. Germain et al, IEEE And Fingerprint Identification Using Delaunay Triangulation G. Bebis et al,IEEE

2 Performance of fingerprint Classification

3 Performance of Classification Techniques (cont..) Confusion Matrix

4 Accuracy Vs Rejection rate

5 Two Databases NIST DB4 - contains 2000 fingerprint pairs NIST DB14 – contains fingerprint pairs –Consist of 8-bit grey level images –Two different fingerprint instances –Classified into 5 classes

6 Results on NIST DB4

7 Results on NIST DB14

8 Accuracy Vs Rejection rate

9 Fingerprint Indexing and Retrieval Problems with classification schemes –Number of classes is small –Fingerprints are unevenly distributed –More than 90% of fingerprints belong to only 3 classes –Difficult to search a single fingerprint form the large database

10 These problems can be handled with 2 different approaches –Fingerprint sub classification –Continuous Classification

11 Fingerprint Sub Classification

12 Continuous Classification and Other Indexing Techniques Uses vectors summarizing their main features Feature vectors are created through a similarity preserving transformation Avoids ambiguous fingerprints System efficiency and accuracy will be balanced by adjusting the size of the neighborhood.

13 Indexing Techniques Using Minutae points Identifies all the minutae triplets in the fingerprints Uses geometric hashing to retrieve a similar fingerprints from the database This is built by quantizing all the possible triplets If the same fingerprint is hit by more triplets, then a voting technique is applied to get the final rank

14 Other Indexing techniques Based on matching scores between the fingerprints In some papers, different Indexing techniques are combined to improve the performance Continuous classification with MKL –based approaches Finger code feature vectors are combined with a simplified version of the minutae triplet approach

15 Retrieval Strategies If exclusive classification is used for indexing then, Hypothesized class only Fixed search order Variable search order

16 If continuous classification is used for indexing then, Fixed radius Incremental search order

17

18 Performance of fingerprint Retrieval

19 Performance of retrieval strategies

20 Performance of retrieval strategies

21 Fingerprint matching using transformation parameter clustering Fingerprint Identification Using Delaunay Triangulation

22 Flash Method Flash algorithm uses a higher a dimensional indexing scheme than geometric hashing by adding invariant properties of the feature subset to the index Second stage uses, transformation parameter clustering to accumulate evidence

23 Flash Method When adding a model to the database, invariant information computed from each subset of feature points forms a key or index Key labels an entry that is added to a multimap, This entry contains the identifier of the model that generated the key and information concerning the feature subset

24

25 When servicing a query, each key generated by the query object is used to retrieve any items in the multimap that are stored under the same index. Each item retrieved represents hypothesized match between subsets of features in the query object and the reference model This hypothesized match is labeled by the reference model by parameters characterizing the geometric transformation bringing the two subsets of features into closest correspondence Votes for these hypothesized matches accumulate in another associative memory structure

26

27 How it applies to fingerprint matching In the fingerprint application, class of transformations that connects different object instances is assumed to be of two-dimensional distance preserving transformations A least squares estimation methodology is used to solve the over constrained pose estimation problem for each hypothesized local correspondence generated by the index lookup process

28 Data abstraction and index generation Minutae provides a natural choice for feature points A triplet of numbers (X, Y, Ө ) represent each feature point

29 Flash matcher uses skeletonized version of the ridge pattern on the finger If a line is drawn between each pair of minutae, the number of ridges crossed by this line can be computed Ridge counting procedure repeats for each pair of minutae in the fingerprint, and the results become part of the flash index

30 The flash algorithm uses redundant combination of three feature points when forming indices This gives some immunity against noise To keep the number of indices generated within bounds, the algorithm restricts the acceptable combinations of feature points used to form an index

31 The search engine requires the generation of indices used for table lookup These indices are descriptive of the objects stored in the database. Each component of the index is invariant under rotations and translations The full index consists of nine components: –Length of each side –Ridge count between each pair –Angles measures with respect to the sides

32 Accumulating evidence During the query phase, each index generated by the query fingerprint This is used to retrieve all the objects in the database that are labeled with same index Each retrieved model objects represents a hypothesized correspondence between 3 points in the query print and three in the model

33 Algorithm that computes the co-ordinate transformation

34 Accumulating evidence If a large number of feature points can be brought into correspondence by rigid transformation of the coordinate system, all of the indices generated by the combinations of three feature points belonging to this set generate the same coordinate transformation parameters

35 Accuracy Issues Four scenarios are possible H 0 is true, and test says H 0 is true H 0 is false, and test says H 0 is true H 1 is true, and test says H 1 is true Two distinct types of errors can be made False Negative: incorrectly assigned mated to non mated False Positive: incorrectly assigned non mated to mated The number of matching triangles that generate a consistent rigid transformation serves as the basis for assigning pairs to the mated or non- mated pair population

36 With the decision criteria, it is straightforward to determine the two error rates from the conditional probability densities computed from the test populations The error rate for incorrectly assigning a mated pair to the nonmated population is given by The error rate for incorrectly assigning a nonmated pair to the mated population is given by

37 Consider one to many identification query The candidate list of hypothesized matches is formed by taking all prints from the reference database Assuming the presence of one mate to the query, the FPR and FNR for and identification search against a database N is shown below The FPR increases drastically with database size because each additional entry in the database provides another opportunity to randomly achieve a high score

38

39 Results Data set inked dab images 657 queries, against this database Query set of prints was a subset of the models They made 657 X comparisons of pairs These pairs divided into 3 groups identical fingerprints(657 pairs) diff. impressions of the same finger( 768 pairs) impressions of different fingers( 64,050,819 pairs)

40

41 Fingerprint identification using Delaunay triangulation

42 Advantages of using this technique Preserves index selectivity Reduces memory requirements Improves recognition time Considers only O(N) minutae triangles

43 Important issues to be consider when using Indexing memory requirements : In the case of fingerprints, memory requirements can become much higher since fingerprints contain more features on the average than typical objects Index selectivity: relates to the discrimination power of the groups considered for indexing groups with low discrimination power give rise to very similar indices large number of hypothetical matches are generated during recognition

44 To deal with this problems Increasing index dimensionality using large size groups Additional information can be computed from each group and added to index

45 Indexing based methods have two phases of operation Preprocessing –features which remain unchanged under geometric transformations are extracted from groups of model points and used to form indices –Indexed locations are filled with entries containing references to the models Recognition –Features from groups of image points are extracted and used to form indices again –The models listed in the indexed entries are collected into a list of candidate models and the most often indexed models are selected for further verification

46 Background on Delaunay Triangulation

47 Delaunay triangulation has certain properties Non degenerate set of points is unique A circle through the three points of a Delaunay triangle contains no other points The minimum angle across all the angles in all the triangles in a delaunay triangulation is greater than the minimum angle in any other triangulation of the same points

48 Indexing using Delaunay Triangulation Minutae triangulation

49 Building the Index Table The index table is built by considering the minutae triangles formed by the Delaunay triangulation From each minutae triangle, information invariant to similarity transformations is computed. Then, an index is formed using the invariants and appropriate information is stored in the indexed table location the Delaunay triangulation, yields O(N).

50 Given a minutae triangle, Compute 3 invariants These based on sides and angles of the triangle First sort the sides of the triangle to avoid considering all possible orders of three points

51 Following invariants are computed

52 After the invariants have been computed followed by quantization yields an integer index The entries stored in the table have the following format

53 Identification step Each index generated by a query fingerprint is used to retrieve all model fingerprints To account for noise, we also retrieve entries stored in a small neighborhood

54 Verification step Performed by aligning the two fingerprints using the transformation computed and by computing the amount of overlap A list of candidate fingerprints which possibly match query fingerprints is generated If a large number of minutae from the candidate fingerprint are close, then it is very likely that the two fingerprints come from the same fingerprint

55 Although we use similarity transformations, differences in the pressure of the finger on the sensor or skin elasticity produce deformations which are not modeled very well by similarity transformations alignment is improved by computing the similarity transformation using affine transformations

56 Experimental results Data set 300 fingerprints, captured from 30 individuals (10 images per finger for each individual) Size is 400 X 400 pixels No restriction on the position and the orientation of fingers

57 Experiments First set of experiment Vary the number of imprints stored in the database for each person Experimented with storing 3, 5,and 7 images per person In each case, 6 experiments were conducted In the first five experiments, images stored in the database are chosen randomly and in the last one, best one is chosen

58 Classify results into 4 categories Correct – query correctly matched to one or more fingerprints from the same person False positive - query matched to one or more fingerprints from the an incorrect person False negative - query has not been matched to any fingerprints from the database Mixed – there is not enough evidence to assign the query fingerprint to one of the previous categories

59 Results

60 Conclusions from the results Recognition accuracy depends on the number of imprints stored in the database for each person Last row of each table shows that if the imprints stored in the database are of good quality, recognition accuracy improved significantly Number of false negatives are relatively high compared to number of false positives

61 Second experiment How false positives increase with the database size Tested how the system performs on fingerprints from people not represented in the database Five experiments were conducted