Fusion of Methods Based on Minutiae, Ridges and Pores for Robust Fingerprint Recognition
FFusion of Methods Based on Minutiae, Ridges and Pores for Robust FingerprintRecognition
Lucas Alexandre RamosUnesp - Brazil Aparecido Nilceu MaranaUnesp - Brazil
Abstract
The use of physical and behavioral characteristics forhuman identification is known as biometrics. Among themany biometrics traits available, the fingerprint is the mostwidely used. The fingerprint identification is based on theimpression patterns, as the pattern of ridges and minutiae,characteristics of first and second levels respectively. Thecurrent identification systems use these two levels of finger-print features due to the low cost of the sensors. However,due the recent advances in sensor technology, it is pos-sible to use third level features present within the ridges,such as the perspiration pores. Recent studies have shownthat the use of third-level features can increase securityand fraud protection in biometric systems, since they aredifficult to reproduce. In addition, recent researches havealso focused on multibiometrics recognition due to its manyadvantages. The goal of this work was to apply fusiontechniques for fingerprint recognition in order to combineminutiae, ridges and pore-based methods and, thus, providemore robust biometrics recognition systems. We evaluatedisotropic-based and adaptive-based automatic pore extrac-tion methods and the fusion of pore-based method with theidentification methods based on minutiae and ridges. Theexperiments were performed on the public database PolyUHRF and showed a reduction of approximately 16% in theEqual Error Rate compared to the best results obtained bythe methods individually.
1. Introduction
Due to the increasing demands for safety and the in-crease in sensors capacity, a great number of studies withthe third-level fingerprint characteristics have been carriedout, as in [4], where the fingerprint fragments recognitionmethod based on ridges proposed by [2] was improved us-ing the sweat pores detected on the ridges in the registra-tion and comparison steps, through its fusion with the ridgesrecognition method. Some fields such as forensics require systems with greatprecision and which are able to adapt to situations wherethe information on the fingerprints are incomplete and onlyfragments are available. According to [5], the use of multi-biometrics systems can significantly improve the accuracyof a system, making it safer and enabling better populationcoverage.Adding pores to a fingerprint based biometrics systemscan make it more secure, since the pores of fingerprints aredifficult to reproduce, and studies indicate they can be usedto distinguish fake silicon fingerprints from real ones, mak-ing the systems safer and more reliable [8].According to [1] the concept of biometrics is defined asthe recognition of individuals based on their physical orbehavioral characteristics, being the fingerprint one of themost commonly used biometric characteristics. The finger-prints presents three levels of characteristics: the patternsformed by macro details of the fingerprint, as the flow of theridges (level one features), the minutiae formed by distinc-tive points observed in the flow of the ridges, as bifurcationsand endings (level two features) and details extracted fromridges, such as the pores and scars (level three features) [1].Figure 1 shows the levels of fingerprint characteristics.Currently only characteristics from level one and two areused on recognition systems [7], [6]. This is due the fact thatthe fingerprint images are generally obtained with resolu-tions up to 500 dpi, which is lower than the resolution pub-lished by NIST (National Institute of Standards and Tech-nology, USA) in 2007 on its guidelines and standards for thethird level features fingerprint recognition. NIST suggeststhe adoption of images with a resolution of at least 1000 dpito allow the proper extraction of these characteristics [9].The minutiae-based systems are the most used ones spe-cially because they employ the method adopted by forensicexperts all over the world and the minutiae is accepted asproof of identity in virtually all countries [10].The goals of this paper are: to implement and evalu-ate fusion techniques for fingerprints recognition methodsbased on minutiae, ridges and pores and to perform themultibiometric recognition using partial fingerprint images. a r X i v : . [ c s . C V ] M a y igure 1. Fingerprint features at Level 1 (upper row), Level 2 (middle row), and Level 3 (lower row).[7]
2. Fingerprint Recognition Method Based onMinutiae
This approach involves the use of a feature vector whoseelements are the descriptors of the minutiae. Therefore,this approachs methods are based on matching points algo-rithms that consists in finding a transformation (translation,rotation, and scale), such that the set of points of the ref-erence image corresponds to the set of points of the queryimage.The minutiae-based recognition method was imple-mented using the biometric systems technology called Ver-iFinger, which is distributed by a company called Neu-rotechnology.VeriFinger uses a set of points (minutiae) for fingerprintrecognition and provides a set of algorithms that increaseperformance and reliability. Some of their characteristicsare, tolerance to deformation, rotation and translation andfilters capable of eliminating noises and improving the qual-ity of deteriorated images. Detailed information about thefingerprint recognition method based on minutiae is not pro-vided by VeriFinger.
This fingerprint recognition method based on ridges pro-posed by Marana and Jain [2], uses the Hough Transformand the feature extraction stage consists of three steps. Firstthe extraction and thinning of ridges based on the algorithmproposed by Jain [11], where the ridges are extracted afterestimating the ridges orientation field and background seg-mentation. The second step is the straight line extraction,where the most significant straight lines from the fingerprintridge pixels are extracted using the Hough Transform. Fi-nally, the last step is the classification of the ridges accord-ing to their lines curvature. Figure 2 illustrates the steps forridge extraction.The comparison consists of two steps, the fingerprintregistration, where the parameters of geometric transforma-tions (rotation, translation and scale) are calculated fromthe Hough Transform using the Hough space peaks, andthe comparison stage where the query image is aligned tothe reference image and the comparison score is calculatedbased on the number of matching ridges between these twoimage using a alignment matrix of M × N dimension where M and N are the number of ridges detected in query imageand the reference image, respectively. igure 2. Ridge Extraction: (a) Fingerprint image; (b) Detectedand thinned ridges from the fingerprint image; (c) Straight linesdetected from a given fingeprint ridge by using the Hough Trans-form Hough Transform [2].
3. Fingerprint Recognition Method Based onRidges and Pores
The recognition method based on ridges and pores pro-posed by [4], aimed to extend the recognition method basedon ridges [2], adding the information from the pores of thefingerprint on the comparison step. The methods used forthe pore extraction step are the following: the method basedon adaptative filters proposed by Zhao [6], and the methodbased on isotropic filters proposed by Ray[13]. Figure 3shows the pores extracted from an image using the adapta-tive filter algorithm and Figure 4 shows the pores extractedusing the isotropic filter algorithm.
Figure 3. Pores extracted from a fingerprint fragment using themethod based on adaptative filters proposed by Zhao [6].Figure 4. Pores extracted from a fingerprint fragment using themethod based on isotropic filters proposed by Ray[13].
In the strategy proposed by Angeloni [4], first the ridge-based method is executed and the transformation parame-ters obtained from the alignment stage are applied to thecoordinates of the pores extracted from the query image,which acts as an additional heuristics in selecting the bestalignment, then the pore comparison step is initiated. The scores obtained by the two methods are fused using aweighted sum, thus, obtaining the final score of the com-parison. Figure 5 shows a diagram of the proposed strategy.
Figure 5. Diagram of the method proposed by Angeloni [4].
In the pore comparison stage a bounding box of 6, 8 and10 pixels is used around each pore extracted due to finger-print natural distortion. Thus, if after the alignment trans-formation a pore is inside the bounding box of the queryimage, then it would be considered a match.
4. The Fingerprint Database
The PolyU HRF [14] is a public database provided by theHong Kong Polytechnic University. The DB PolyU HRFI dataset consists of 1480 images, 10 images collected for148 people in two separate sections. Images were capturedusing a high-resolution sensor, with resolution of 1200 dpiand dimension of 240x320 pixels.The PolyU HRF database is quite challenging becausein addition to presenting common problems like non-lineardistortion, rotation and translation, the area of comparisonof the fingerprint fragments is very small. Figure 6 showssome fingerprints fragments found in this database.The experiments were performed according to the pro-tocol used by Zhao [12], which consists of two types ofcomparisons, namely:Genuine comparisons: The database is composed of fivesamples of fingerprints of each individual, obtained in twodifferent sessions, in a total of 10 images per individual. Forthe genuine fingerprint comparisons, each fingerprint imageof the second session was compared with all images of thesame fingerprint of the first session, totaling 3700 genuinecomparisons.Impostor comparisons: For the impostor fingerprintcomparison tests the first fingerprint of each finger from thesecond session was compared with the first image from firstsession of all the other fingerprints totaling 21756 impostorcomparisons. igure 6. Examples of fingerprints fragments found in the PolyUHRF database.
5. Experimental Results
In the experiments we analyzed the recognition accuracyusing the following combinations: i) only minutiae; ii) onlyridges; iii) only pores; iv) minutiae and ridges; v) ridges andpores; vi) minutiae and pores; and vii) minutiae, ridges andpores. For the results evaluation the Equal Error Rate (EER)was calculated for each experiment refered previously. Allthe methods were fused using a score-level fusion techniqueusing weights for each score. The final weights were cho-sen on all possible combinations. Table 5 shows the individ-ual results obtained by each method before fusion. Table 5shows the results obtained for the fusion in pais (ridges andpores, ridges and minutiae and finally pores and minutiae).Method EERMinutiae based 25,08%Ridges based 23,50%Pores Based (Isotropic) 26,02%Pores Based (Adaptative) 23,22%
Table 1. EER obtained for each method individually.
Fused Methods EERRidges and Pores Isotropic 22,01%Ridges and Pores Adaptative 22,31%Minutiae and Ridges 9,35%Minutiae and Pores (Isotropic) 10,45%Minutiae and Pores (Adaptative) 9,08%
Table 2. Fusion of methods in pairs (minutiae, ridges and pores).
Finally, the fusion of the three methods was performed.The weights used in the weighted sums were obtained usingall possible combinations. Table 5 shows the best results forthe three methods fusion.Fused Methods EERMinutiae, Ridges and Pores( Isotropic) 8,57%Minutiae, Ridges and Pores (Adaptative) 8,74%
Table 3. Fusion of methods based on Minutiae, Ridges and Pores
It’s important to stress that the best results were obtainedwith a higher weight assigned to the recognition methodbased on minutiae.
6. Conclusion
The recognition of fingerprint fragments proved to bequite challenging, even when using several methods, and itis of great importance in many areas, particularly forensics,where fingerprints fragments are common and their identi-fication is essential.The great accuracy improvement is due to the lowamount of minutiae found in the fingerprints fragments,while the pores and ridges are abundant and may cause mis-classification errors, minutiaes can be used as a confirma-tion whether the fingerprint if genuine or impostor, reducingtherefore the false accpetance rate. The recognition methodbased on minutiae proved to be very accurate in identifingimpostor fingerprint fragments, however it was not very ef-ficient in identifying genuine ones, proving that commercialsystems also have difficulty in recognizing fingerprint frag-ments and may benefit from the fusion of multiple finger-print characteristics .The recognition methods based on ridges and pores pre-sented more balanced results, with both showing problemsin identifying genuine and impostors fingerprint fragments,and are seldom used commercially, mainly because theirstudy is recent, and its sensors are highly expensive.The fusion of the methods proved to be quite effec-tive, decreasing by more than 16% the EER when minutiae,ridges and pores were fused making the system more reli-able. The study of multibiometrics fusion proved to be verypromising and relevant, since the fingerprint is the mostwidely used biometric characteristics used for identifing in-dividuals.
7. Acknowledgement
The authors would like to thank Fapesp for the financialsupport. eferenceseferences