Fingerprint RecognitionFingerprint Recognition Anil K. Jain Michigan State University jain@cse.msu.edu http://biometrics.cse.msu.edu © Jain, 2004 OutlineOutline • Brief History • Fingerprint Representation • Minutiae-based Fingerprint Recognition • Fingerprint Enhancement • Other methods of Fingerprint Recognition • Ridge Feature Maps • Correlation • Fingerprint Classification • Fingerprint Individuality • Estimating non-linear deformation © Jain, 2004 FingerprintsFingerprints • Description: graphical flow like ridges present in human fingers • Formation: during embryonic development • Permanence: minute details do not change over time • Uniqueness: believed to be unique to each finger • History: used in forensics and has been extensively studied © Jain, 2004 Biological Principles of FingerprintsBiological Principles of Fingerprints • Individual epidermal ridges and valleys have different characteristics for different fingerprints • Configurations and minute details of individual ridges and valleys are permanent and unchanging (except for enlargement in the course of bodily growth) • Configuration types are individually variable, but they vary within limits which allow for systematic classification © Jain, 2004 © Jain, 2004 History of FingerprintsHistory of Fingerprints Fingerprint on Palestinian lamp (400 A.D.) Bewick’s trademark (1809) A Chinese deed of sale (1839) signed with a fingerprint © Jain, 2004 History of FingerprintsHistory of Fingerprints • Many impressions of fingers have been found on ancient pottery • Grew (1684): first scientific paper on ridges, valleys & pore structures • Mayer (1788): detailed description of anatomical formation of fingerprints • Bewick (1809): used his fingerprint as his trademark • Purkinje (1823): classified fingerprints into 9 categories based on ridges • Herschel (1858): used fingerprints on legal contracts in Bengal • Fauld (1880): suggested "scientific identification of criminals" using fingerprints • Vucetich (1888): first known user of dactylograms (inked fingerprints) • Scotland Yard (1900): adopted Henry/Galton system of classification • FBI (1924) set up a fingerprint identification division with a database of 810,000 fingerprints • FBI (1965): installed AFIS with a database of 810,000 fingerprints • FBI (2000): installed IAFIS with a database of 47 million 10 prints; conducts an average of 50,000 searches/ day; ~15% of searches are in lights out mode. Response time: 2 hours for criminal search and 24 hours for civilian search © Jain, 2004 Fingerprint MatchingFingerprint Matching • Find the similarity between two fingerprints Fingerprints from the same finger Fingerprints from two different fingers © Jain, 2004 Fingerprint ClassificationFingerprint Classification • Assign fingerprints into one of pre-specified types Plain Arch Tented Arch Right Loop Left Loop Plain WhorlAccidental Pocket Whorl Double Loop © Jain, 2004 © Jain, 2004 Fingerprint SensorsFingerprint Sensors • Optical, capacitive, ultrasound, pressure, thermal, electric field © Jain, 2004 Types of Fingerprint ImagesTypes of Fingerprint Images Flat Fingerprint (One-touch print from a single-finger livescan device) Unsegmented Slap Fingerprint (4-finger simultaneous impression from livescan devices or scanned from paper FP cards) Rolled Fingerprint (Image collected by rolling the finger across the livescan platen or paper from nail to nail) © Jain, 2004 TerminologyTerminology • Fingerprint – Impression of a finger • Minutiae – Ridge bifurcations,endings and many other features (52 types listed, 7 are usually used by human experts and two by automated systems) • Core – uppermost point on the innermost ridge • Delta – separating point between pattern area and non-pattern area • Pattern class – determined by ridge flow characteristics © Jain, 2004 Fingerprint RepresentationFingerprint Representation • Local ridge characteristics (minutiae): ridge ending and ridge bifurcation • Singular points: Discontinuity in ridge orientation Core Delta Ridge Bifurcation Ridge Ending © Jain, 2004 Minutiae-based RepresentationMinutiae-based Representation © Jain, 2004 Challenges in Fingerprint IdentificationChallenges in Fingerprint Identification • Cuts and bruises on finger; dry or oily fingers; ~3% of fingerprints are not of “good” quality • Wear and tear of sensor; no proven contactless fingerprint sensor technology is currently available • New compact solid-state sensors capture only a small portion of the fingerprint • Fingerprint impression is often left on the sensor • Non-universality of fingerprint • Changes in sensor technology; sensor interoperability Non-universality of fingerprint Fake fingerprint © Jain, 2004 Fingerprint Matching TechniquesFingerprint Matching Techniques • Minutiae-based • Correlation-based • Ridge Feature-based © Jain, 2004 Minutiae-based Fingerprint Verification System Minutiae-based Fingerprint Verification System © Jain, 2004 Steps in Minutiae ExtractionSteps in Minutiae Extraction • Orientation field estimation • Fingerprint area location • Ridge extraction • Thinning • Minutia extraction © Jain, 2004 Minutiae Extraction AlgorithmMinutiae Extraction Algorithm © Jain, 2004 Minutiae Extraction AlgorithmMinutiae Extraction Algorithm Ridge FilterOrientation EstimationInput Image Minutiae Extraction Post-processing Ridge Thinning Minutiae Detection © Jain, 2004 Orientation FieldOrientation Field • Angle formed by the ridges with horizontal axis • Usually computed at the block level and not at the pixel level (more resistant to noise) • The gradient direction is perpendicular to the ridge direction • An arithmetic average of gradient angles is not well defined • Doubling the angles allows local gradient estimates to be averaged © Jain, 2004 Orientation Field Estimation AlgorithmOrientation Field Estimation Algorithm • Divide the input fingerprint image into nonoverlapping blocks of size W x W • Compute the gradients Gx (i, j) and Gy (i, j) at each pixel (i, j) using Sobel or Marr-Hildreth operator • Least squares estimate of the local orientation of the block centered at (i, j) is ⎟ ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎜ ⎝ ⎛ − ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ = ∑ ∑ + −= + −= − 2 2 2 2 22 1 ),(),( ),(),(2 tan 2 1 ),( W i W iu W j W jv yx yx vuGvuG vuGvuG jiθ © Jain, 2004 Orientation Field Estimation AlgorithmOrientation Field Estimation Algorithm • Local ridge orientation varies slowly in a neighborhood where no singular points appear • A low-pass filter can be applied to modify the local ridge orientation • To apply low pass filter the orientation image is converted into a continuous vector field )),(2sin(),( )),,(2cos(),( jiji andjiji y x θ θ =Φ =Φ © Jain, 2004 Orientation Field Estimation AlgorithmOrientation Field Estimation Algorithm • Apply a 2-D low pass filter as follows where is the size of the filter, default is 5 x 5 blocksΦΦ × ww ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ Φ Φ ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ = − ),( ),( tan 2 1 ),( ' ' 1 ji ji jiO y x Local ridge orientation at (i, j) is © Jain, 2004 Orientation Field ConsistencyOrientation Field Consistency • Consistency (actually inconsistency) level of the orientation field in the local neighborhood is If C (i, j) is greater than Tc, re-estimate the local orientations in this block at a lower resolution level, until C (i, j) is less than Tc © Jain, 2004 Fingerprint Region LocalizationFingerprint Region Localization • Separation of the fingerprint area from the background • In the fingerprint area, the variance is very high in a direction orthogonal to the ridge orientation • In the background regions, the variance has no directional dependence • Local variance of gray level can be used to locate the fingerprint area • Assumption: Only one fingerprint is present in the image © Jain, 2004 Fingerprint LocationFingerprint Location Certainty level is defined as , ),( ),(),(1 ),( 2 22 2 ⎟ ⎟ ⎠ ⎞ ⎜ ⎜ ⎝ ⎛ + ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ = jiG jiGjiG W ji e yx ε ( ).),(),(),( 2 2 2 2 22 ∑ ∑ + −= + −= += w i W iu W j W jv yxe vuGvuGjiG For a block, if the certainty level is below a threshold Tk, then all pixels in that block are marked as background © Jain, 2004 Ridge ExtractionRidge Extraction • Ridges and furrows run parallel to one another forming 2-D sine wave • Gray-level values on the ridges attain their local maxima along a direction that is orthogonal to the local orientation • The fingerprint image is convolved with two masks ht (i, j; u, v) and hb (i, j ; u, v) that are 180º out of phase © Jain, 2004 Ridge Extraction FilterRidge Extraction Filter © Jain, 2004 Ridge ExtractionRidge Extraction • On average, L x H is 11 x 7. Ideally the mask width must be equal to the width of the local ridge • Pixel (i, j) is labeled as ridge pixel if both the graylevel values of the pixel (i, j) of the convolved images are greater than Tridge • The masks also do some smoothing, determined the parameter δ • For computational efficiency, δ is made large enough to make the filter coefficient degenerate to a single value © Jain, 2004 Removing Holes and SpecklesRemoving Holes and Speckles • The binary ridge map has many holes and speckles due to noise, smudges, breaks, etc. • Connected component algorithm is used to remove holes and speckles • Remove the connected components having less than 50 pixels • Thinning is applied to obtain 8-connected thinned ridges of one pixel width © Jain, 2004 Minutiae DetectionMinutiae Detection • Spikes and breaks in the thinned ridge map leads to detection of spurious minutiae • Heuristics are applied to post-process the thinned ridge map • If the angle between the branch and the trunk is greater than 70º and less than 110º, then the branch is removed • If the break in the ridges is less than 15 pixels and no other pixel passes through it, then the break is connected © Jain, 2004 Minutiae DetectionMinutiae Detection © Jain, 2004 Minutiae Type DetectionMinutiae Type Detection • A ridge pixel is a ridge ending, if the number of ridge pixels in the 8-neighborhood is 1 • A ridge pixel is a ridge bifurcation, if the number of ridge pixels in the 8-neighborhood is greater than or equal to 3 • A ridge pixel is a intermediate ridge pixel, if the number of ridge pixels in the 8-neighborhood is 2 • [x, y, θ, associated ridge] are stored for each minutia © Jain, 2004 Minutiae Point Pattern MatchingMinutiae Point Pattern Matching © Jain, 2004 Minutiae MatchingMinutiae Matching • Point pattern matching problem • Let ( ) ( ){ }P M P M P M PPP yxyxP θθ ,,,...,,, 111= be the set of M minutiae in the template image • Let ( ) ( ){ }Q N Q N Q N QQQ yxyxQ θθ ,,,...,,, 111= be the set of N minutiae in the input image • Find the number of corresponding minutia pairs between P and Q and compare it against a threshold © Jain, 2004 Stages in Minutiae MatchingStages in Minutiae Matching • Alignment – Estimate rotation, translation, and distortion – Input fingerprint is aligned with the template • Matching – Compute the similarity between the pre-aligned input and the template using the number of matching minutiae © Jain, 2004 Minutiae MatchingMinutiae Matching Generate Alignment Hypothesis Find Similarity Using String Matching Template Matching score Input © Jain, 2004 Alignment of point patternsAlignment of point patterns • Corresponding point pairs are not known • Due to noise and deformations, the input minutiae cannot be aligned exactly with respect to the template minutiae • Ridges corresponding to the minutia are aligned © Jain, 2004 Alignment of Input and Template Ridges Alignment of Input and Template Ridges © Jain, 2004 Alignment AlgorithmAlignment Algorithm • Ridges are represented as 1-D discrete signals, normalized wrt the inter-ridge distance • Let Rd and RD be the set of ridges in the input and template images • Each ridge d є Rd is matched with D є RD as L is the minimal length of the two ridges, di and Di represent the distances from point i to the x-axis © Jain, 2004 Alignment AlgorithmAlignment Algorithm • If S is greater than a threshold (0.8), go to the next step. Otherwise continue to match the next pair of ridges • Compute the translation vector and rotation angle as follows: γi, and Γi are the radial angles of the ith point with respect to the reference minutia © Jain, 2004 Alignment AlgorithmAlignment Algorithm • Let (xd, yd, θd)T be the reference minutia (based on which the transformation parameters are estimated). The N input minutiae are transformed as follows: © Jain, 2004 Matching AlgorithmMatching Algorithm • Estimate the rotation and translation parameters and align the two minutiae patterns • Convert the template and input patterns into the polar coordinate system with respect to the reference minutiae and represent them as strings by concatenating each minutia in an increasing order of radial angles ( ) ( ){ }P M P M P M PPP p ererP θθ ,,,...,,, 111= ( ) ( ){ }Q N Q N Q N QQQ p ererQ θθ ,,,...,,, 111= © Jain, 2004 String FormationString Formation © Jain, 2004 String Matching • Match the strings Pp and Qp with a modified dynamic programming algorithm to find the edit distance • Edit distance must have the elastic property of string matching • Edit distance is defined recursively as follows: © Jain, 2004 © Jain, 2004 String MatchingString Matching • α, β, and γ are the weights associated with radius, radial angle, and orientation components • δ, ε, and ρ specify the bounding box • Ω is the pre-specified penalty for mismatch • This scheme can tolerate non-linear deformations and inexact transformations to some extent, but cannot compensate for it • An adaptive bounding box can be used © Jain, 2004 Bounding Box AdjustmentBounding Box Adjustment © Jain, 2004 © Jain, 2004 String MatchingString Matching • δl(m,n), δh(m,n), εl(m,n), and εh(m,n) specify the adaptive bounding box; initial values depend on the resolution of the fingerprint image • η is the learning rate • Parameter values are determined empirically © Jain, 2004 Minutiae CorrespondencesMinutiae Correspondences © Jain, 2004 Matching ScoreMatching Score • The edit distance is used to establish the correspondence of the minutiae between Pp and Qp • Total number of corresponding minutiae is computed as MPQ • The matching score is MN MM S PQPQ100 = © Jain, 2004 Minutiae Matching ResultMinutiae Matching Result (a) (c) (b) (d) © Jain, 2004 2-D Dynamic Programming based Minutiae Matching 2-D Dynamic Programming based Minutiae Matching Matching Stage Alignment Stage Maximize L Template Minutiae T = {t1, .., tm} Query Minutiae Q = {q1, .., qn} For each selected minutiae pair (ti, qj), compute ∆x, ∆y, ∆θ 1-D elastic string matching based on ti, qj, ∆x, ∆y, ∆θ 2-D dynamic programming based matching using (ti, qj, ,∆x, ∆y, ∆θ )Lmax (ti, qj, ∆x, ∆y, ∆θ)Lmax maxL′ Minutiae Pairing S1(R(ti),R(qj)) > 1δ S2(O(T),O(Q),∆x, ∆y, ∆θ ) > 2δ Score / (S2,L’max) Find an alignment that maximizes the number of minutiae correspondences S1 → Ridge similarity measure, S2 → Orientation similarity measure, R(t) → 1-D representation of ridge points of minutia t, O(T) → Orientation field Matching time ~0.1 sec. © Jain, 2004 Matching Score DistributionsMatching Score Distributions • Performance depends on the database. FVC2002 Database 1 (100 users, 8 impressions/user) • For FAR = 0.1% (1 in 1000), GAR = 97.1% • EER = 1.65%; at 0% False Accept, FRR = 4% Matching Score Distribution ROC Curve © Jain, 2004 Analysis of ErrorsAnalysis of Errors • Minutiae Extraction – Extraction stage does not extract all minutiae and their ridges – There may be no corresponding minutiae having ridge points for finding the correct alignment • Alignment – Corresponding minutiae with ridge points exist – Alignment step fails due to small number of correspondences • Matching – Estimated alignment is correct – But, the matching score is low because the number of correspondences is low compared to the number of minutiae – Reasons: deformation, spurious and missing minutiae © Jain, 2004 Minutiae Extraction FailureMinutiae Extraction Failure A B True Minutiae Matches: A1→B3, A18→B9, A19→B7 A1, B9 and B7 were detected, but the associated ridges were not detected because they are close to the boundary © Jain, 2004 Alignment FailureAlignment Failure BA True Minutiae Matches: A7→B9, A8→B8, A4→B1 A7→B9 and A8→B8 pairs have ridge points; however, there exists a false alignment that results in more than three matches © Jain, 2004 Matching FailureMatching Failure BA No. of matching minutiae identified by the matcher = 10 No. of minutiae in A = 38; No. of minutiae in B = 34 Spurious minutiae and large deformation leads to small score