Fingerprint Alignment Using Special Ridges - Semantic Scholar
Fingerprint Alignment Using Special Ridges Chunfeng Hu, Jianping Yin, En Zhu, Hui Chen, Yong Li School of Computer Science, National University of Defense Technology, Changsha, PR China [email protected] Abstract Fingerprint is one of the biometrics used to identify a person. Alignment is an important step in fingerprint recognition, affecting greatly the speed and accuracy of matching. Eight types of Special ridges are introduced to align two fingerprints. The ridge with maximum of sampled curvature is used as reference ridges for initial alignment. And corresponding special ridges paired by topology get aligned by their features. The alignment parameters of translation and rotation finally come from all aligned special ridge pairs. Experiments show that alignment using special ridges is fast and robust.
1. Introduction Fingerprint is one of deeply researched and widely used biometrics to identify a person, because of its universality, distinctiveness, permanence and easy collectability. The typical fingerprint identification procedure consists of fingerprint image acquisition, feature extraction and fingerprint matching. The innovations of fingerprint sensors have basically satisfied the demands of increasing applications of Automated Fingerprint Identification System (AFIS) now. However feature extraction of low quality images and matching of badly nonlinear deformed fingerprints are not thoroughly solved. Alignment of input and template fingerprints is an important step just after features extracted and before matching. The purpose of alignment is to estimate the translation and rotation parameters between input and template fingerprints, so as to get the overlap of the two fingerprints for matching. The algorithms of alignment affect greatly the speed and accuracy of matching, and misalignment of two fingerprints of the same finger results in false matching. Different alignment algorithms are based on different features extracted. Characteristic fingerprint features are generally described as global and local levels.
978-1-4244-2175-6/08/$25.00 ©2008 IEEE
Global features are the macro details of the whole image, such as ridge flow and pattern type, orientation field and so on. Local features are those of a point, ridge or block, which are only from a part of the fingerprint, such as minutia. The singular points (core and delta), which indicate the ridge flow and pattern types, are prominent enough to align fingerprints directly. Nilsson [1] detected the core point by complex filters applied to the orientation field in multiple resolution scales, and the translation and rotation parameters are simply computed by comparing the coordinates and orientation of the two core points. Jain [2] predefined four types of kernel curves: arch, left loop, right loop and whorl, each with several subclasses respectively. These kernel curves were fitted with the image, and then used for alignment. Yager [3] proposed a two stage optimization alignment combined both global and local features. It first aligned two fingerprints by orientation field, curvature maps and ridge frequency maps, and then optimized by minutiae. The alignment using global features is fast but not robust, because the fingerprints from the same finger may be the impressions of different area of the fingertip, and some missing and spurious local features make the global features different. Local features are always robust but need more computation in alignment. Minutiae features contain most of a fingerprint’s individuality, and are the most important features for fingerprint verification system. Ratha [4] proposed the Generalized Hough Transform for alignment. The translation and rotation of each pair of potentially matching minutiae need be calculated and the most likely translation and rotation parameters were chosen by the minutiae pair so-called reference minutiae. Algorithms of alignment based on minutiae always need test all possible minutiae pairs. They are accurate but computationally expensive. And many spurious minutiae are extracted from low quality images. It increases the computation, and may also result in false alignment and matching. Several other features are combined with minutiae for alignment. Jiang [5] exploited the structural information of
distances between minutiae, relative differences between minutiae directions, minutiae types and ridge counts to find a likely correspondence minutiae pair. Feng [6] combined minutiae-based and texture-based descriptors of every minutia, and chose the top n most similar minutiae pairs for alignment. But choosing the reference minutiae directly induces that the adjacent minutiae pairs align well but the minutiae pairs far away do not align satisfactorily. Optimization of the alignment of reference minutiae to find a transformation which minimized the distance between all corresponding minutiae pairs was proposed by Ramoser [7]. Zhu [8] introduced global alignment of all corresponding minutiae pairs after selecting the reference minutiae. They made great improvement in accuracy of alignment, but did not take into account of the prominent features to reduce the computation. In this paper we use special ridges to align the input and template fingerprints. Compared with other features, special ridges are more reliable than singular points, and prominent enough to align directly. Eight types of special ridges extracted from thinned image are introduced in section 2. In section 3 the alignment using these special ridges is described. Section 4 shows the experimental results, and the conclusions and future works are presented in section 5.
three neighborhoods on the ridge, while there is only one for termination and plain end point. Second, plain end point is next to the background while termination not. The ridge stops at a termination or plain end point, and a new ridge begins tracing. The branch ridges are to be traced using Deep First Search algorithm at a bifurcation. Curvature is a measure of the rate of change of tangent along the ridge. The curve is represented using arc with three sampled points at a fixed length on ridge, and the curvature at every sampled point is calculated, illustrated in Figure 1. The ridge tracing finishes when another end point encounters. After tracing a ridge, the maximum of the curvature and coordinates and orientation of this sample point are recorded as MaxC, and the length as l. And points on the ridge are labeled by the ridge ID, as gray value storing in image. At last the two end points reorder clockwise, and ridge type is defined by the combination of two end points. Special ridges are described in section 2.2, and others are common ridges. After tracing the whole foreground, all ridges and end points get IDs from 1 respectively, which are no more than 255 in our experiment. Some ridges and their end points are removed by ridge length, because they are usually fragment due to low quality area of image.
2. Special Ridges Extraction A fingerprint image consists of alternative black and white curves, which are the impression of ridges and valleys on fingertip. Supposing the fingerprint image has been thinned after preprocessing of orientation estimation, segmentation and enhancement using circular gabor filter [9] and so on, the ridges and minutiae are extracted while tracing the thinned image.
2.1. Ridge Tracing A Depth First Search algorithm is used to trace ridges in the thinned image. The ridge features, such as end points, curvature at sample points, and length are extracted on every ridge. The first ridge usually starts with an end point. Besides two types of traditional minutiae, termination and bifurcation, a new type of end point is extracted, which is on the boundary between foreground and background. These plain end points interrupt by the reason of limitation area of acquired image, and always differ due to different impression of the fingertip. These three types of end points are detected by the neighborhood in thinned image and position in segmented image. First, bifurcation is that there are
Figure 1. extracting features of ridges
A topology of ridges is constructed at all end points. Now R0 is the current ridge (Figure 1). At the plain end point or termination, supposing a line perpendicular to its orientation, it intersects with two adjacent ridges, which are defined as inside ridge R1 and outside ridge R2 respectively. At the bifurcation, according to the angles between the three lines from the bifurcation to the sample points on each ridge, two branches, R1 and R2 are the edges of the smallest angle, and the other ridge R0 is the parent ridge. Three ridges reorder clockwise from the parent ridge. At last, the ridge set R and end point set P extracted are denoted as
R = {Psi , Pei , li , MaxCi , RTi }iN=r1
(1)
P = {xi , yi , θi , R 0i , R1i , R 2i , MT }
Np i i =1
(2)
where Nr, Np are the numbers of ridges and points respectively, Psi , Pei are the start and end point of the
ith ridge by clockwise order, RTi is the type of ith ridge, xi, yi, θi, MTi are the x- and y-coordinates, orientation and type of ith point.
2.2. Special Ridges Eight types of special ridges (Figure 2) are defined by the types of its endpoints curvature and topology. These special ridges are not only described as common ridges in R, but also reserve all sampled points on ridges and some other features, such as the angle α of two lines from the sampled maximum curvature point to two end points.
Figure 2. special ridges
(1) RT_MAXC: the maximum of sampled curvature of the ridge is the largest of all. This ridge is usually the most inner one and the point that the largest curvature is at is always near the core point. Of course the core point may be on its adjacent ridges because of some ridges interrupt in the area of core point. (2) RT_INSELF: the inside ridge at one end point is itself. This type of ridge is usually in the area of core point, and maybe also RT_MAXC. (3) RT_INOUT: the inside and outside ridges of one end point are the same ridge. It is usually one short ridge in area of core point, or one of broken ridges near RT_MAXC. (4) RT_2BIFUR: both end points of the ridge are bifurcations. It does not include the parent ridge divaricates into two branches and then joins again, which is called RT_RING. Its length is a reliable feature in alignment. (5) RT_NOTIN: two branches at a bifurcation are not adjacent. The bifurcation is near the delta point. The angle between two branches is useful for alignment. (6) RT_HELIX: the ridge turns around itself. This fingerprint is usually a whorl. The length is not robust in different instance of a fingerprint, because
one of end points varies due to different area acquired. The ridge may turn clockwise or counter-clockwise. And the number of ridges and minutiae between RT_HELIX are reliable for alignment too. (7) RT_RING: the parent ridge divaricates into two branches and then joins again. The two branches are both defined as RT_RING. The length is as useful as that of RT_2BIFUR. (8) RT_CIRCLE: it is a circle with no end points. The parameters of end points are invalid. It may represent as an ellipse, and the perimeter, major axis and minor axis are used to depict it.
3. Fingerprint Alignment Every fingerprint should certainly have a ridge of RT_MAXC, so it is chosen as the initial ridge for alignment, which is called reference ridges. There are two core points in fingerprint of whorl, and curvatures are computed at sampled points, so the RT_MAXC point may appear a little away form core point, even on different ridges. So several largest MaxC are chosen as candidates of RT_MAXC, and the adjacent ridges are also used as potential reference ridges. RT_MAXC, RT_INSELF, RT_INOUT are always related to core point, and may be the potential reference ridges too. The initial translation and rotation parameters are computed by the coordinates and orientation of the point of maximum curvature. As for other special ridges, corresponding ridges are found by the topology based on the reference ridges. Some special ridges can be corresponding to other types, even common ridges, while RT_NOTIN, RT_HELIX and RT_CIRCLE can only be corresponding to the same type of special ridges. Every corresponding ridge pair is aligned with different features (Table 1), by the difference between each feature compared to given threshold. Table 1. features for alignment Special ridge types Features for alignment RT_MAXC MaxC, α RT_INSELF MaxC, α RT_INOUT MaxC, α RT_2BIFUR MaxC, α, l RT_NOTIN* MaxC, angles of 3 ridges RT_HELIX* MaxC, ridge count RT_RING l, end points RT_CIRCLE* l, major axis, minor axis
Usually not all eight types of special ridges exist in one fingerprint image, and different instances of a finger may not have the same ones. There are no more than seven special ridges of a fingerprint and two
fingerprints with more than two aligned special ridge pairs are regarded to be aligned in our experiments. At last, the mean values of translation and rotation parameters of each aligned special ridges pairs are those of the global alignment.
4. Experimental Results The FVC2006 fingerprint database, DB3 set A is used in the experiments[10]. There are 1,680 images (140 fingers, 12 images from every finger) are captured by thermal sweeping sensor, with the resolution of 500 dpi [11]. We took part in FVC2006 as P053 anonymously. That algorithm achieved 25th in DB3 and 18th in average four databases in open category. We compare two algorisms with same method of feature extracting and matching, but different alignment using special ridges instead of trying all minutiae pairs. Matching results show in Table 2. The alignment using special ridges (SPR) improves the accuracy of matching. It is because that the special ridges are prominent and accurate. Accuracy improvement comes from better alignment of fingerprints with similar minutiae but different ridges. But since the ridges are extracted from thinned image, the matching results are relative to not only the alignment algorithm, but also the features extracted. A great number of low quality images still result in false matching. Another reason is that the method proposed still does not take into account of nonlinear deformation. Table 2. matching on FVC 2006 DB3_A
EER FMR100 FMR1000 ZeroFMR ZeroFNMR
P053 5.74% 9.84% 13.99% 18.00% 100.00%
SPR 3.87% 6.93% 9.25% 14.89% 100%
5. Conclusions A fingerprint is the pattern of ridges and valleys on the surface of a fingertip. Ridges are the intuitional features according to human observation. Eight types of special ridges introduced in this paper are prominent and reliable to align two fingerprints fast and robustly. The ridge with the maximum of sampled curvature is used for initial alignment. Other corresponding ridges then align using different features. Experiments show great improvement compared to alignment only using minutiae.
Since enhancement and thinning introduce some spurious features, including minutiae and ridges. We are considering correcting ridges in original gray image after tracing from the thinned image. And the ridges could also be used for matching, especially small image with little minutiae. We are also considering matching each special ridge pairs to get the parameters of deformation in local area, which may be quite effective to match fingerprints with nonlinear deformation. It is a promising problem to represent and verify fingerprints by ridges.
Acknowledgement This work is supported by National Natural Science Foundation of China under grant NO.60603015.
References [1] K. Nilsson and J. Bigun, "Prominent symmetry points as landmarks in fingerprint images for alignment," in Proceedings - International Conference on Pattern Recognition, 2002, pp. 395-398. [2] A. K. Jain and S. Minut, "Hierarchical kernel fitting for fingerprint classification and alignment," in Proceedings - International Conference on Pattern Recognition, 2002, pp. 469-473. [3] N. Yager and A. Amin, "Fingerprint alignment using a two stage optimization," Pattern Recognition Letters, 2006, vol. 27, pp. 317-324. [4] N. K. Ratha, "A real-time matching system for large fingerprint databases," IEEE Transactions on Pattern Analysis and Machine Intelligence, 1996, vol. 18, pp. 799-813. [5] X. Jiang and W.Y. Yau, "Fingerprint Minutiae Matching Based on the Local and Global Structures," in Proceedings - International Conference on Pattern Recognition, 2000, pp. 1042-1045. [6] J. Feng, "Combining minutiae descriptors for fingerprint matching," Pattern Recognition, 2008, vol. 41, pp. 342352. [7] H. Ramoser, B. Wachmann, and H. Bischof, "Efficient alignment of fingerprint images," in Proceedings International Conference on Pattern Recognition, 2002, pp. 748-751. [8] E. Zhu, J. Yin, and G. Zhang, "Fingerprint matching based on global alignment of multiple reference minutiae," Pattern Recognition, 2005, vol. 38, pp. 16851694. [9] E. Zhu, J. Yin, and G. Zhang, "Fingerprint Enhancement Using Circular Gabor Filter," in Image Analysis and Recognition, 2004, pp. 750-758. [10] J. Fierrez, J. Ortega-Garcia, D. Torre Toledano, and J. Gonzalez-Rodriguez, "Biosec baseline corpus: A multimodal biometric database," Pattern Recognition, 2007, vol. 40, pp. 1389-1392. [11] FVC, website of Fingerprint Verification competition 2006, url: http://bias.csr.unibo.it/fvc2006/.
1. Introduction Fingerprint is one of deeply researched and widely used biometrics to identify a person, because of its universality, distinctiveness, permanence and easy collectability. The typical fingerprint identification procedure consists of fingerprint image acquisition, feature extraction and fingerprint matching. The innovations of fingerprint sensors have basically satisfied the demands of increasing applications of Automated Fingerprint Identification System (AFIS) now. However feature extraction of low quality images and matching of badly nonlinear deformed fingerprints are not thoroughly solved. Alignment of input and template fingerprints is an important step just after features extracted and before matching. The purpose of alignment is to estimate the translation and rotation parameters between input and template fingerprints, so as to get the overlap of the two fingerprints for matching. The algorithms of alignment affect greatly the speed and accuracy of matching, and misalignment of two fingerprints of the same finger results in false matching. Different alignment algorithms are based on different features extracted. Characteristic fingerprint features are generally described as global and local levels.
978-1-4244-2175-6/08/$25.00 ©2008 IEEE
Global features are the macro details of the whole image, such as ridge flow and pattern type, orientation field and so on. Local features are those of a point, ridge or block, which are only from a part of the fingerprint, such as minutia. The singular points (core and delta), which indicate the ridge flow and pattern types, are prominent enough to align fingerprints directly. Nilsson [1] detected the core point by complex filters applied to the orientation field in multiple resolution scales, and the translation and rotation parameters are simply computed by comparing the coordinates and orientation of the two core points. Jain [2] predefined four types of kernel curves: arch, left loop, right loop and whorl, each with several subclasses respectively. These kernel curves were fitted with the image, and then used for alignment. Yager [3] proposed a two stage optimization alignment combined both global and local features. It first aligned two fingerprints by orientation field, curvature maps and ridge frequency maps, and then optimized by minutiae. The alignment using global features is fast but not robust, because the fingerprints from the same finger may be the impressions of different area of the fingertip, and some missing and spurious local features make the global features different. Local features are always robust but need more computation in alignment. Minutiae features contain most of a fingerprint’s individuality, and are the most important features for fingerprint verification system. Ratha [4] proposed the Generalized Hough Transform for alignment. The translation and rotation of each pair of potentially matching minutiae need be calculated and the most likely translation and rotation parameters were chosen by the minutiae pair so-called reference minutiae. Algorithms of alignment based on minutiae always need test all possible minutiae pairs. They are accurate but computationally expensive. And many spurious minutiae are extracted from low quality images. It increases the computation, and may also result in false alignment and matching. Several other features are combined with minutiae for alignment. Jiang [5] exploited the structural information of
distances between minutiae, relative differences between minutiae directions, minutiae types and ridge counts to find a likely correspondence minutiae pair. Feng [6] combined minutiae-based and texture-based descriptors of every minutia, and chose the top n most similar minutiae pairs for alignment. But choosing the reference minutiae directly induces that the adjacent minutiae pairs align well but the minutiae pairs far away do not align satisfactorily. Optimization of the alignment of reference minutiae to find a transformation which minimized the distance between all corresponding minutiae pairs was proposed by Ramoser [7]. Zhu [8] introduced global alignment of all corresponding minutiae pairs after selecting the reference minutiae. They made great improvement in accuracy of alignment, but did not take into account of the prominent features to reduce the computation. In this paper we use special ridges to align the input and template fingerprints. Compared with other features, special ridges are more reliable than singular points, and prominent enough to align directly. Eight types of special ridges extracted from thinned image are introduced in section 2. In section 3 the alignment using these special ridges is described. Section 4 shows the experimental results, and the conclusions and future works are presented in section 5.
three neighborhoods on the ridge, while there is only one for termination and plain end point. Second, plain end point is next to the background while termination not. The ridge stops at a termination or plain end point, and a new ridge begins tracing. The branch ridges are to be traced using Deep First Search algorithm at a bifurcation. Curvature is a measure of the rate of change of tangent along the ridge. The curve is represented using arc with three sampled points at a fixed length on ridge, and the curvature at every sampled point is calculated, illustrated in Figure 1. The ridge tracing finishes when another end point encounters. After tracing a ridge, the maximum of the curvature and coordinates and orientation of this sample point are recorded as MaxC, and the length as l. And points on the ridge are labeled by the ridge ID, as gray value storing in image. At last the two end points reorder clockwise, and ridge type is defined by the combination of two end points. Special ridges are described in section 2.2, and others are common ridges. After tracing the whole foreground, all ridges and end points get IDs from 1 respectively, which are no more than 255 in our experiment. Some ridges and their end points are removed by ridge length, because they are usually fragment due to low quality area of image.
2. Special Ridges Extraction A fingerprint image consists of alternative black and white curves, which are the impression of ridges and valleys on fingertip. Supposing the fingerprint image has been thinned after preprocessing of orientation estimation, segmentation and enhancement using circular gabor filter [9] and so on, the ridges and minutiae are extracted while tracing the thinned image.
2.1. Ridge Tracing A Depth First Search algorithm is used to trace ridges in the thinned image. The ridge features, such as end points, curvature at sample points, and length are extracted on every ridge. The first ridge usually starts with an end point. Besides two types of traditional minutiae, termination and bifurcation, a new type of end point is extracted, which is on the boundary between foreground and background. These plain end points interrupt by the reason of limitation area of acquired image, and always differ due to different impression of the fingertip. These three types of end points are detected by the neighborhood in thinned image and position in segmented image. First, bifurcation is that there are
Figure 1. extracting features of ridges
A topology of ridges is constructed at all end points. Now R0 is the current ridge (Figure 1). At the plain end point or termination, supposing a line perpendicular to its orientation, it intersects with two adjacent ridges, which are defined as inside ridge R1 and outside ridge R2 respectively. At the bifurcation, according to the angles between the three lines from the bifurcation to the sample points on each ridge, two branches, R1 and R2 are the edges of the smallest angle, and the other ridge R0 is the parent ridge. Three ridges reorder clockwise from the parent ridge. At last, the ridge set R and end point set P extracted are denoted as
R = {Psi , Pei , li , MaxCi , RTi }iN=r1
(1)
P = {xi , yi , θi , R 0i , R1i , R 2i , MT }
Np i i =1
(2)
where Nr, Np are the numbers of ridges and points respectively, Psi , Pei are the start and end point of the
ith ridge by clockwise order, RTi is the type of ith ridge, xi, yi, θi, MTi are the x- and y-coordinates, orientation and type of ith point.
2.2. Special Ridges Eight types of special ridges (Figure 2) are defined by the types of its endpoints curvature and topology. These special ridges are not only described as common ridges in R, but also reserve all sampled points on ridges and some other features, such as the angle α of two lines from the sampled maximum curvature point to two end points.
Figure 2. special ridges
(1) RT_MAXC: the maximum of sampled curvature of the ridge is the largest of all. This ridge is usually the most inner one and the point that the largest curvature is at is always near the core point. Of course the core point may be on its adjacent ridges because of some ridges interrupt in the area of core point. (2) RT_INSELF: the inside ridge at one end point is itself. This type of ridge is usually in the area of core point, and maybe also RT_MAXC. (3) RT_INOUT: the inside and outside ridges of one end point are the same ridge. It is usually one short ridge in area of core point, or one of broken ridges near RT_MAXC. (4) RT_2BIFUR: both end points of the ridge are bifurcations. It does not include the parent ridge divaricates into two branches and then joins again, which is called RT_RING. Its length is a reliable feature in alignment. (5) RT_NOTIN: two branches at a bifurcation are not adjacent. The bifurcation is near the delta point. The angle between two branches is useful for alignment. (6) RT_HELIX: the ridge turns around itself. This fingerprint is usually a whorl. The length is not robust in different instance of a fingerprint, because
one of end points varies due to different area acquired. The ridge may turn clockwise or counter-clockwise. And the number of ridges and minutiae between RT_HELIX are reliable for alignment too. (7) RT_RING: the parent ridge divaricates into two branches and then joins again. The two branches are both defined as RT_RING. The length is as useful as that of RT_2BIFUR. (8) RT_CIRCLE: it is a circle with no end points. The parameters of end points are invalid. It may represent as an ellipse, and the perimeter, major axis and minor axis are used to depict it.
3. Fingerprint Alignment Every fingerprint should certainly have a ridge of RT_MAXC, so it is chosen as the initial ridge for alignment, which is called reference ridges. There are two core points in fingerprint of whorl, and curvatures are computed at sampled points, so the RT_MAXC point may appear a little away form core point, even on different ridges. So several largest MaxC are chosen as candidates of RT_MAXC, and the adjacent ridges are also used as potential reference ridges. RT_MAXC, RT_INSELF, RT_INOUT are always related to core point, and may be the potential reference ridges too. The initial translation and rotation parameters are computed by the coordinates and orientation of the point of maximum curvature. As for other special ridges, corresponding ridges are found by the topology based on the reference ridges. Some special ridges can be corresponding to other types, even common ridges, while RT_NOTIN, RT_HELIX and RT_CIRCLE can only be corresponding to the same type of special ridges. Every corresponding ridge pair is aligned with different features (Table 1), by the difference between each feature compared to given threshold. Table 1. features for alignment Special ridge types Features for alignment RT_MAXC MaxC, α RT_INSELF MaxC, α RT_INOUT MaxC, α RT_2BIFUR MaxC, α, l RT_NOTIN* MaxC, angles of 3 ridges RT_HELIX* MaxC, ridge count RT_RING l, end points RT_CIRCLE* l, major axis, minor axis
Usually not all eight types of special ridges exist in one fingerprint image, and different instances of a finger may not have the same ones. There are no more than seven special ridges of a fingerprint and two
fingerprints with more than two aligned special ridge pairs are regarded to be aligned in our experiments. At last, the mean values of translation and rotation parameters of each aligned special ridges pairs are those of the global alignment.
4. Experimental Results The FVC2006 fingerprint database, DB3 set A is used in the experiments[10]. There are 1,680 images (140 fingers, 12 images from every finger) are captured by thermal sweeping sensor, with the resolution of 500 dpi [11]. We took part in FVC2006 as P053 anonymously. That algorithm achieved 25th in DB3 and 18th in average four databases in open category. We compare two algorisms with same method of feature extracting and matching, but different alignment using special ridges instead of trying all minutiae pairs. Matching results show in Table 2. The alignment using special ridges (SPR) improves the accuracy of matching. It is because that the special ridges are prominent and accurate. Accuracy improvement comes from better alignment of fingerprints with similar minutiae but different ridges. But since the ridges are extracted from thinned image, the matching results are relative to not only the alignment algorithm, but also the features extracted. A great number of low quality images still result in false matching. Another reason is that the method proposed still does not take into account of nonlinear deformation. Table 2. matching on FVC 2006 DB3_A
EER FMR100 FMR1000 ZeroFMR ZeroFNMR
P053 5.74% 9.84% 13.99% 18.00% 100.00%
SPR 3.87% 6.93% 9.25% 14.89% 100%
5. Conclusions A fingerprint is the pattern of ridges and valleys on the surface of a fingertip. Ridges are the intuitional features according to human observation. Eight types of special ridges introduced in this paper are prominent and reliable to align two fingerprints fast and robustly. The ridge with the maximum of sampled curvature is used for initial alignment. Other corresponding ridges then align using different features. Experiments show great improvement compared to alignment only using minutiae.
Since enhancement and thinning introduce some spurious features, including minutiae and ridges. We are considering correcting ridges in original gray image after tracing from the thinned image. And the ridges could also be used for matching, especially small image with little minutiae. We are also considering matching each special ridge pairs to get the parameters of deformation in local area, which may be quite effective to match fingerprints with nonlinear deformation. It is a promising problem to represent and verify fingerprints by ridges.
Acknowledgement This work is supported by National Natural Science Foundation of China under grant NO.60603015.
References [1] K. Nilsson and J. Bigun, "Prominent symmetry points as landmarks in fingerprint images for alignment," in Proceedings - International Conference on Pattern Recognition, 2002, pp. 395-398. [2] A. K. Jain and S. Minut, "Hierarchical kernel fitting for fingerprint classification and alignment," in Proceedings - International Conference on Pattern Recognition, 2002, pp. 469-473. [3] N. Yager and A. Amin, "Fingerprint alignment using a two stage optimization," Pattern Recognition Letters, 2006, vol. 27, pp. 317-324. [4] N. K. Ratha, "A real-time matching system for large fingerprint databases," IEEE Transactions on Pattern Analysis and Machine Intelligence, 1996, vol. 18, pp. 799-813. [5] X. Jiang and W.Y. Yau, "Fingerprint Minutiae Matching Based on the Local and Global Structures," in Proceedings - International Conference on Pattern Recognition, 2000, pp. 1042-1045. [6] J. Feng, "Combining minutiae descriptors for fingerprint matching," Pattern Recognition, 2008, vol. 41, pp. 342352. [7] H. Ramoser, B. Wachmann, and H. Bischof, "Efficient alignment of fingerprint images," in Proceedings International Conference on Pattern Recognition, 2002, pp. 748-751. [8] E. Zhu, J. Yin, and G. Zhang, "Fingerprint matching based on global alignment of multiple reference minutiae," Pattern Recognition, 2005, vol. 38, pp. 16851694. [9] E. Zhu, J. Yin, and G. Zhang, "Fingerprint Enhancement Using Circular Gabor Filter," in Image Analysis and Recognition, 2004, pp. 750-758. [10] J. Fierrez, J. Ortega-Garcia, D. Torre Toledano, and J. Gonzalez-Rodriguez, "Biosec baseline corpus: A multimodal biometric database," Pattern Recognition, 2007, vol. 40, pp. 1389-1392. [11] FVC, website of Fingerprint Verification competition 2006, url: http://bias.csr.unibo.it/fvc2006/.
