Presentation on theme: "ONLINE FINGERPRINT VERIFICATION"— Presentation transcript:
1 ONLINE FINGERPRINT VERIFICATION Sharat ChikkerurCenter for Unified Biometrics and SensorsUniversity at BuffaloAdvisor: A. N. CartwrightCommittee: V. Govindaraju, A. H. Titus, L. Kondi
2 Abstract Background Challenges Contributions Traditional password/token based authentication schemes are insecure and are being replaced by biometric authentication mechanismsFingerprints were one of the first biometrics to be widely usedDespite 40 years of research, fingerprint recognition is still an open problem.ChallengesFeature extraction is very unreliable in poor quality printsMatching fingerprints under non linear distortion is difficultContributionsNew fingerprint image enhancement using STFT analysis.New feature extraction algorithm based on chain code contoursNew graph based matching algorithm robust to non linear distortion
4 Biometrics Definition Examples Physical Biometrics Biometrics is the science of verifying and establishing the identity of an individual through physiological features or behavioral traitsExamplesPhysical BiometricsFingerprint, Hand Geometry, FaceMeasurement BiometricDependent on environment/interactionBehavioral BiometricsHandwriting, Signature, Speech, GaitPerformance/Temporal biometricDependent on state of mindChemical/Biological BiometricsSkin spectroscopyDNA, blood-glucoseBiometrics offers a promising solution for reliable and uniform identification and verification of an individual. Biometrics is the science of verifying and establishing the identity of an individual through physiological features or behavioral traits. Physical biometrics rely on physiological features such as fingerprints, hand geometry, iris pattern, facial features etc. for identity verification. Behavioral biometrics depends upon behavioral features such as speech patterns, handwriting, signature, walking gait etc. for authentication. These traits are unique to an individual and hence cannot be misused, lost or stolen.Physiological biometrics are more stable and robust than behavioral biometrics and a single sample is enough to obtain identifying information. Biometrics such as signature and speech are acquiredand learnt over time and also subject to the user’s state of mind and disposition. Multiple samples are required to acquire a stable representation of user features.Biometrics are based on established scientific principles as a basis for authentication.
5 Fingerprints as a Biometric High UniversalityA majority of the population (>96%) have legible fingerprintsMore than the number of people who possess passports, license and IDsHigh DistinctivenessEven identical twins have different fingerprints (most biometrics fail)Individuality of fingerprints established through empirical evidenceHigh PermanenceFingerprints are formed in the fetal stage and remain structurally unchanged through out life.High PerformanceOne of the most accurate forms of biometrics availableBest trade off between convenience and securityHigh AcceptabilityFingerprint acquisition is non intrusive. Requires no training.The skin of the fingers is corrugated by a pattern of contours. These contour ridges are formed during foetal development and remain unchanged throughout a person’s life. It has been established that fingerprint patterns are unique to each individual. Of all the biometric techniques, fingerprint based authentication has been the most well established and well researched topic. It has been used in forensic science for authentication since the early 1900s.Fingerprints have been an important tool used by for law enforcement and forensics for over a century. Automatic Fingerprint Identification Systems (AFIS) can provide absolute identification of an individual by processing the image of a fingerprint. Fingerprints are formed while the foetus is 4 months old and remain unchanged through out an individual’s life time
6 Fingerprints 101: Fingerprint Classes A fingerprint is made up of system of oriented friction ridgesA fingerprint can be classified based on type the ridge flow patternThe corrugated surface of the fingerprint is made up of ridges and valleys cover that the entire palmer surface of the hand. The flow pattern of these ridges and valleys are unique to each individual. These patterns that are used for identification and authentication. The image below shows the image of a fingerprint along with the distinguishing features on the print.The flow of the ridges and patterns has been classified into 5 broad classes. This classification is used to catalog the fingerprints and also in authenticating two prints. Henry systems follow an elaborate classification scheme of cataloging and filing forensic prints. Fig1 shows the different classes of ridge flows. This methods cannot be used to distinguish between two fingerprints.Fig 2 shows the distinguishing features on the fingerprint. These features are discontinuities or anomalies in the normal flow of ridges on the surface of the finger. These features are termed as minutiae (small details). There are eighteen different types of minutiae. Fig1 shows the most commonly encountered ones and their names. Fig 2 shows a thumbprint captured on paper. These are typically the kind of images that forensic AFIS (Automatic Fingerprint Identification Systems) have to deal with. The quality of the print not only deteriorates during capture but also during storage and hence AFIS systems are more sophisticated than their biometric counterparts.Classification helps in narrowing down possible matchesIn reality, the class distribution is skewed (>65% are loops)Used only in law enforcement applications
7 Fingerprints 101: Ridge Characteristics Fingerprints can be distinguished based on the ridge characteristicsRidge characteristics mark local discontinuities in the ridge flowNo two individuals have the same pattern of ridge characteristics at the same relative locationsLocal FeaturesGlobal Features
8 Prior Related Work: Matching Paradigms ManualHuman experts use a combination of visual, textural, minutiae cues and experience for verificationStill used in the final stages of law enforcement applicationsImage basedUtilizes only visual appearance.Requires the complete image to be stored (large template sizes)Texture basedTreats the fingerprint as an oriented texture imageLess accurate than minutiae based matchers since most regions in the fingerprints carry low textural contentMinutiae basedUses the relative position of the minutiae pointsThe most popular and accurate approach for verificationResembles manual approach very closely.
9 Image Based Matching: Optical Correlation AdvantagesImage itself is used as the templateRequires only low resolution imagesOptical correlation makes it extremely fast (Choudary and Awwal ’99, Lee et al. 99, Roberge et al. 99, Baze et al.00)DisadvantagesImage itself is used as the template (template size about 30 KB)Requires accurate alignment of the two prints (unreliable in poor prints)Not robust to changes in scale, orientation and position.
10 Texture Based Matching: Filterbanks AdvantagesUses texture information (lost in optical and minutiae based schemes)Performs well with poor quality printsFeatures are statistically independent from minutiae and can be combined with minutiae matchers for higher accuracy (Jain et al. 00, Jain et al 01)DisadvantagesRequires accurate alignment of the two prints (unreliable in poor prints)Not invariant to translation, orientation and non-linear distortion.Less Accurate than minutiae based matchers
11 Minutiae Based Matching AdvantagesInvariant to translation, rotation and scale changesVery accurate (Ratha et al 96, Jain et al. 97, Jian Yau 00, Bazen and Garez 03)DisadvantagesMinutiae extraction is error prone is low quality imagesNot robust to non-linear distortion.Does not use visual and textural cues
13 Outline Introduction Fingerprint Image Enhancement Need for EnhancementPrior Related WorkProposed Algorithm: STFT AnalysisExperimental EvaluationMinutia Feature ExtractionMatching AlgorithmSoftware Demos
14 Need for EnhancementWhat you seeWhat you ‘think’ you see
15 Reality: What you usually get.. High contrast printTypical dry printFaint printThe performance of any fingerprint recognizer depends upon the quality of the fingerprint image processed. AFIS systems have achieved high performances in case of reasonably good prints. However there is not yet satisfactory methods to deal with bad quality fingerprints as often encountered in forensic and biometric applications. Effective methodologies for cleaning the valleys between the ridge contours are lacking in current fingerprint recognition systems. The figures show some prints acquired under different conditions. It can be seen that the minutiae features cannot be easily extracted from the original gray scale image. Therefore some form on enhancement is required before processing the fingerprint image further.Low contrast printTypical Wet PrintCreases
16 Challenges Challenges Fingerprint image is non stationary (has dominant local orientation and frequency)General purpose image processing algorithms are not usefulTraditional operators and filters assume Gaussian noise model‘Noise’ in fingerprint images consists mostly of ridge breaksContextual FiltersExisting techniques are based on ‘contextual’ filteringFilter parameters are adapted to each local neighborhoodFilter parameters in ‘unrecoverable’ regions can be interpolated based on its neighbors
17 Prior Related Work: Spatial Filtering (Yang et.al 1996, Greenberg et. Al 1999) proposed local anisotropic filteringFilter kernel adapts at each pixel locationHong et al, 96/98 proposed the use of Gabor filters for enhancementGabor filter has the best joint space-frequency localizationDoes not handle high curvature regions well due to block wise approach.Even Symmetric KernelFourier spectrum showing the localization
18 Prior Related Work: Fourier Domain Filtering Sherlock et al 94, proposed the use of Fourier domain filteringThe image is convolved with a filter bank of directionally selective filtersImage enhanced by selecting a linear combination of filter responsesWatson et al. 94, proposed the use or ‘root filtering’ for enhancement.(Pseudo matched filter)Does not require the computation of orientation imagesRoot FilteringFourier Domain Filtering
19 Traditional Approaches Local Orientation(x,y)Gradient MethodEnhancementFrequency/SpatialThe fingerprint can be seen as an oriented texture. This property of the fingeprints can be used while enhancing the fingerprint image. Traditional image processing methods such as gaussian smoothening, or low pass filtering cannot be used as they tend to bridge the gaps between the ridges. The ridges have to be enhanced only in a direction parallel to their orientation. There are spatial and frequency domain methods that perform such kind of filtering. Spatial domain methods are based on anisotropic filters aligned in the ridge direction(??) or are based on Gabor filtering(Anil Jain et al). Frequency domain methods utilizing FFT (Grother et al, Monroe et al) are also present and are more successful. The enhancement depends upon the accurate estimation of the local ridge orientation and local ridge spacing within the image. Presently the local orientation and the local ridge spacing are obtained through separate algorithms. The enhancement is performed using this information. This approach requires multiple passes and is computationally expensive. We propose a unified approach to fingerprint image enhancement using FFT analysis. The proposed method extracts the local orientation and ridge spacing information in one pass and at the same time performs the enhancement.Local Ridge SpacingF(x,y)Projection Based Method[Ratha et al 95]
20 Proposed Approach: Overview RegionMaskSTFTAnalysisFrequencyImageFourier domainEnhancementThe proposed methods uses FFT based frequency domain analysis to estimate the ridge spacing and orientation in the image. The enhancement is also done in the frequency domain. The analysis yields the energy map, orientation map and ridge spacing map corresponding to the image. The energy map indicates the presence of ridges and their contrast within the image. The energy map can be successfully used to segment the fingerprint image from the background. The orientation map provides the orientation of the ridges in a local neighborhood. This information is used in the enhancement procedure. The ridge spacing map provides the inter ridge distance variation in the image. The ridge spacing can be used in the enhancement procedure to design a suitable band pass filter that allows ridges and eliminates noise and gray level gradients across the image.The following slides show the results of the FFT analysis algorithm.OrientationImageCoherenceImage
21 STFT AnalysisFingerprint image is non stationary, so we require both space and frequency resolution: time frequency analysisSTFT in 1DSTFT in 2D
22 Surface Wave ModelFingerprint ridges can be modeled as an oriented waveLocal ridge orientationLocal ridge frequencySurface waveLocal NeighborhoodsValidity of the model
23 Parameter EstimationParadigm: The Fourier domain response can be viewed as a distribution of surface waves. Each term F(r, θ) corresponds to a surface wave of frequency 1/r and orientation θWe seek to find the most likely surface wave and hence estimate the dominant direction and frequencyWe can represent the Fourier spectrum in polar form as F(r,θ) The power spectrum is reduced to a joint probability density function usingThe angular and frequency densities are given by marginal density functionsThe enhancement techniques can be divided into two distinct classes (I) Spatial domain techniques and (ii) Frequency domain based techniquesSpatial domain techniques: This technique involves convolving a filter kernel(mask) with the entire image to obtain the filtered image. Standard spatial domain filters such as isotropic gaussian filters and other low pass filters cannot be applied to fingerprint images as they blur the image in all directions. Fingerprint image is a highly oriented structure. The spatial filter that is used for enhancement should blur or average the pixels in the direction of the ridge at the same time increasing the contrast between the ridges and valleys in a direction perpendicular to the ridge. In general, spatial domain filters are computationally less expensive compared to frequency domain based filters.Frequency domain filters are based on enhancement in the Fourier domain. The most common approaches include the block level filtering as proposed by Grother, Candela et al. In this method the Fourier transform of the block is multiplied by its power spectrum raised to a fractional power. This is similar to matched filtering approach found in communication and signal processing systems. Other approaches such as Morno et al. are based on directionally selective filters. The image is passed through a filter bank where each filter is oriented in a given direction. The resulting image is obtained by combining the images obtained from each of the directional filter. In general, Fourier based methods are computationally expensive but yield better results when compared to spatial filters.Most of the existing methods in fingerprint image enhancement are based on ‘Contextual filtering’, where the filter parameters adapt to the local ridge orientation and frequency.Anisotropic filters are a special case of spatial contextual filters. Unlike standard gaussian kernel, the anisotropic filter kernel is not uniform in all directions. It is designed to smooth selectively in the direction of the ridge. The shape of the kernel adapts by using the local ridge orientation and ridge frequency. The advantage of this approach is that it creates far fewer spurious ridges and artifacts compared to Fourier based methods and is more robust to noise.Gabor filters can be applied in both spatial and frequency domain. Unlike other directional filters, Gabor filters are both directionally and frequency selective filters. It has been shown that the Gabor kernel has the optimal space frequency resolution. In this approach, the image is divided into blocks and the orientation and ridge frequency is estimated in each block. The block is then filtered using a Gabor kernel that has the corresponding orientation and frequency. The final image is obtained by combining the filtered block images.
27 Coherence ImageBlock processing is unreliable in regions of high curvatureSherlock and Monro 94, relax filter parameters near the singular locationsEstimation of singular point is difficult in poor images!We use an angular coherence measure proposed by Rao and Jain 90
28 EnhancementThe slide shows the results of the frequency domain enhancement. Presently the enhancement approach suggested by Grother et al is used to perform the enhancement. The frequency and orientation map obtained in the previous section is not included in the filter design. The results will only improve with their inclusion.
31 Qualitative Comparison(cont.) Gabor Filter based EnhancementProposed Approach
32 Objective EvaluationWe evaluated the effect of enhancement on 800 images from FVC2002 DB3The evaluation consists of 2800 genuine test and 4950 impostor testsIt can be seen that the matcher performance improves with enhancement
34 Background Minutiae represent local discontinuities in ridge flow Minutiae features are the most widely used fingerprint representationThere are several standards such as CBEFF (file format) and ANSI-NIST (interchange format) standards for minutiae based fingerprint representationMinutiae extraction approaches may be broadly categorized intoBinarization based approachesDirect gray scale extraction
35 Prior Related Work Binarization Approaches MINDTCT,NIST NFIS, (Garris et. Al, 02)Directionally adaptive binarizationTemplate matching is used to detect minutiaeAdaptive Flow Orientation technique (Ratha et. al., 95)Binarization is performed by peak detectionPeak detection leads to false positives in regions of poor ridge constrast.Direct Gray Scale Ridge FollowingRidge Following (Maio and Maltoni 97, Jiang and Yau 01)Based on ridge pursuitHas low computational complexity.Cannot handle poor contrast prints and images with poor ridge structure.Relies on a good orientation map for ridge pursuitThe enhancement techniques can be divided into two distinct classes (I) Spatial domain techniques and (ii) Frequency domain based techniquesSpatial domain techniques: This technique involves convolving a filter kernel(mask) with the entire image to obtain the filtered image. Standard spatial domain filters such as isotropic gaussian filters and other low pass filters cannot be applied to fingerprint images as they blur the image in all directions. Fingerprint image is a highly oriented structure. The spatial filter that is used for enhancement should blur or average the pixels in the direction of the ridge at the same time increasing the contrast between the ridges and valleys in a direction perpendicular to the ridge. In general, spatial domain filters are computationally less expensive compared to frequency domain based filters.Frequency domain filters are based on enhancement in the Fourier domain. The most common approaches include the block level filtering as proposed by Grother, Candela et al. In this method the Fourier transform of the block is multiplied by its power spectrum raised to a fractional power. This is similar to matched filtering approach found in communication and signal processing systems. Other approaches such as Morno et al. are based on directionally selective filters. The image is passed through a filter bank where each filter is oriented in a given direction. The resulting image is obtained by combining the images obtained from each of the directional filter. In general, Fourier based methods are computationally expensive but yield better results when compared to spatial filters.Most of the existing methods in fingerprint image enhancement are based on ‘Contextual filtering’, where the filter parameters adapt to the local ridge orientation and frequency.Anisotropic filters are a special case of spatial contextual filters. Unlike standard gaussian kernel, the anisotropic filter kernel is not uniform in all directions. It is designed to smooth selectively in the direction of the ridge. The shape of the kernel adapts by using the local ridge orientation and ridge frequency. The advantage of this approach is that it creates far fewer spurious ridges and artifacts compared to Fourier based methods and is more robust to noise.Gabor filters can be applied in both spatial and frequency domain. Unlike other directional filters, Gabor filters are both directionally and frequency selective filters. It has been shown that the Gabor kernel has the optimal space frequency resolution. In this approach, the image is divided into blocks and the orientation and ridge frequency is estimated in each block. The block is then filtered using a Gabor kernel that has the corresponding orientation and frequency. The final image is obtained by combining the filtered block images.
37 Proposed Approach: Chain Code Contours Provides a lossless description of the contour and also gives direction and curvature information.Translation and rotation invariantUsed in computer vision for encoding object boundariesUsed for character recognition (Madhavanth et. al 99)
38 Minutiae Detection using Chain Codes Minutiae are encountered as points of ‘significant’ turn on the contourLeft turn: Ridge endingRight turn: Bifurcation
42 Experimental Evaluation Test Data150 prints from FVC2002(DB1) were randomly selected for evaluation.Ground truth was established using a semi automated truthing tool.Results compared using NIST NFIS open source software.MetricsProposed by Sherlock et. Al 94Sensitivity: Ability of the algorithm to detect true minutiaeSpecificity : Ability of the algorithm to avoid false positivesFlipped : Minutiae whose type has been exchanged
48 Challenges: Quality and Intra-user variance Variation in qualityIntra-user variation
49 Prior Related Work :Global Matching Point correspondences not known : combinatorial problemRelaxation approach (Ranade and Rosenfield 93)Likelihood of each match is either decreased or increased at each iteration based on compatibility of rest of the pointsIterative approach makes it too slow to be practicalGeneralized Hough Transform (Ratha et al. 96)All possible transformation represented as a quantized search spaceSearches for the most optimal transform in the search spaceVery fastRidge Alignment (Jain et al. 97)Performs explicit alignment before matchingEach minutiae is associated with its ridge (represented by a curve)The alignment is based on ridge correspondenceGlobal matching is then performed using string edit distance
50 Prior Related Work: Local Matching (Jiang and Yau 00)11 dimensional local features derived from reference minutiae and two closest neighborsBest match is used only for explicit alignment(Jea and Govindaraju 04)5 dimesional features Si (ri0, ri1, φi0, φi1, δi) derived from two closest neighborsAlignment is still required(Ratha et al. 00)‘Star’ representation derived from all minutiae within a particular radiusConsolidation by checking consistency(Garris et. al 03: BOZORTH3)Line featuresConsolidation by linking consisting matchesInstead of the 3-elements minutiae representation, we use this 5-elements secondary feature representation.For a given feature point or called central minutia, we find the two nearest neighboring minutiae of it. The secondary feature contains the information of the distances between the central minutia to the neighboring minutiae, the orientation between them and the angle between this two line segments.The secondary feature is derived only from the widely-used minutiae representation, thus our algorithm can be easily adapted other systems.We do not use the ridge count information, because it is not required information in minutiae representation.We do not use the minutiae types as the feature, because they are not reliable.Here is an example, these two feature points are extracted from the same finger but in different impressions. One is interpreted as a bifurcation while the other one is seen as a ridge ending.Secondary feature also has some other advantages.It is a pure localized feature. It doesn’t rely on any global landmarks.It is orientation invariant. There is no pre-alignment stage needed.
51 Proposed Algorithm Representation K-Plet Features invariant to rotation and translationLocal relationship formally represented by a directed graphLocal MatchingPosed as a string alignment problem and solved by dynamic programmingMatches all neighbors simultaneouslyConsolidationCoupled Breadth First SearchBreadth first search is used to propagate the matchesSimilar to human verificationNo explicit alignment required at any stage
54 Local MatchingAll local neighbors have to be matched simultaneously. Greedy approach does not work when conflicts occurThese can solved by finding the alignment through optimization process such as by solving a string alignment problemExample of alignment:S (acbcdb) – (ac__bcdb)T (cadbd) - (_cadb_d_)Trivial solution requires exponential timeEach match is given a cost. Alignment solved through recurrence relation
61 Important Differences Traditional Breadth First SearchTraversal Defined only over a single graphAll neighbors are considered for expanding the pathCoupled Breadth First SearchTraversal proceeds in two directed graphs simultaneouslyOnly ‘matched’ neighbors are considered for expanding the pathConstant number of neighbors provides a bound for the traversal complexity
62 Experimental Evaluation 800 prints from FVC2002(DB1)2800 genuine tests,4950 impostor testsCompared with BOZORTH 3Error RatesBOZORTH3: 3.6% EER, 5.0% FMR100Proposed: 1.5% EER, 1.65% FMR100
63 Software CUBS Truthing Tool CUBS Minutiae Truthing Tool CUBS Fingerprint Verification DemoMatlab code for Fingerprint EnhancementMatlab Toolbox for Fingerprint Verification1179 downloads since 01/30637 downloads since 03/22
64 Conclusion Contributions New Fingerprint Image Enhancement using STFT Analysis.Simultaneously estimates all intrinsic imagesIncreases recognition rate of existing matchersNew Feature Extraction Algorithm using Chain code ContourObviates need for thinningPerforms favorably with NIST feature extractorNew Graph based matching algorithmRobust to non linear distortionFormal technique for propagating local matchesPerforms better than NIST BOZORTH3 matcher over FVC DB1 database
65 Acknowledgements Tsai Yang Jea (Alan) Chaohang Wu Sergey Tulyakov Faisal FarooqAmit MhatreKarthik SridharanSankalp NayakRest of the research group at CUBS