Algorithm for Detection and Elimination of False Minutiae in ...
Algorithm for Detection and Elimination of False Minutiae in Fingerprint Images Seonjoo Kim, Dongjae Lee, and Jaihie Kim Department of Electrical and Electronics Engineering,Yonsei University, Seoul, Korea [email protected]
Abstract. A common problem in fingerprint recognition is the existence of false minutiae which increase both FAR and FRR in fingerprint matching. In this paper, a robust minutiae postprocessing algorithm is proposed. Unlike most algorithms which use simple distance and connectivity criterions for postprocessing, we also used orientation and flow of ridges as the key factor for postprocessing to avoid eliminating true minutiae while postprocessing. It is shown by the experiments that our postprocessing algorithm improves the minutiae extraction accuracy and the performance of the matching process.
1
Introduction
Most fingerprint recognition systems are based on minutiae matching [1]. Minutiae are local discontinuities of fingerprints and are restricted to two types : ridge ending and ridge bifurcation [2], [3]. A common problem in fingerprint recognition is the existence of false minutiae which increase both FAR and FRR in fingerprint matching. Therefore, the enhancement of the fingerprint image and the false minutiae elimination form an important part of the system. However, most of the researches emphasized on the fingerprint image enhancement and the false minutiae elimination process was based on simple distance and connectivity criteria [1], [2], [3], [4]. But the problem with such simple approaches is that it eliminates true minutiae while eliminating false minutiae. Xiao and Raafat proposed in [5], a minutiae postprocessing algorithm based on both statistical and structural information. However, their method relies heavily on connectivity which makes it complex and unreliable to bad quality fingerprints. Also, specific structural informations were not given. In this paper, we present an efficient minutiae postprocessing algorithm. The goal of our minutiae extraction algorithm is to remove as many false minutiae as possible while retaining true minutiae. The goal is achieved by postprocessing minutiae based on not only the minutiae distance and connectivity but also using the orientation and flow of ridges as the key factor. Rest of the paper is organized as follows. Section 2 briefly decribes the adopted preprocessing procedures. Section 3 decribes the proposed minutiae postprocessing procedures. The performance of the proposed algorithm is shown by experiments in Section 4. Finally, Section 5 contains conclusion.
2 2.1
Preprocessing and Minutiae Extraction Preprocessing
Preprocessing procedures necessary for minutiae extraction are shown in Fig.1.
Fig. 1. Preprocessing Procedures
The first preprocessing procedure is the calculation of the local ridge orientation. The least mean square orientation estimation algorithm[6] is used and the local ridge orientation is specified by blocks rather than every pixel. The calculated orientation is in the range between 0 and π After the ridge orientation calculation, ridge frequency is calculated[6]. Using the calculated orientations and frequencies, the input grayscale image is enhanced and binarized by Gabor filters which have both frequency-selective and orientation-selective properties[6]. The final preprocessing operation required before extracting minutiae is thinning. Thinning reduces the the widths of the binary ridges down to a single pixel to facilitate the job of detecting ridge endings and bifurcations. The Zhang-Seun thinning algorithm[7] is used in this paper. 2.2
Minutiae Extraction
After a thinned fingerprint image is obtained, minutiae are directly extracted from the thinned image. To detect minutiae, a count of the pixel value transition at a point of interest in a 3×3 mask is used [4], [5]. If the count equals 2, then the point is an endpoint. If the count equals 6, then the point is a bifurcation. For each extracted minutia, the x & y coordinate and the orientation are recorded. The minutiae orientation is defined as the local ridge orientation of the associated ridge[2]. The minutiae orientation is in the range between 0 and π.
3
Proposed Minutiae Postprocessing Algorithm
There are many false minutiae among the extracted minutiae. False minutiae will decrease the performance of the fingerprint identification system by increasing both FRR and FAR. Typical false minutiae structures are shown in Fig.2. In this section, minutiae postprocessing algorithm is proposed. To eliminate false minutiae without eliminating true minutiae, the proposed methods are based on the flow of ridges as well as the minutiae distance and connectivity.
(a)
(b)
(c)
(d)
(e)
(f)
(g)
Fig. 2. Typical false minutiae : (a)Broken ridge, (b)Bridge, (c)Short ridge, (d)Short ridge, (e) Short Ridge, (f) Hole, (g)Triangle
(a)
(b)
Fig. 3. False Minutiae Structures: (a)Broken ridge, (b)Bridge
3.1
Detecting Broken Ridge Structure
Because of scars and insufficient finger pressure on the input device, a ridge may break into two ridges creating two endpoints. Obviously, these two endpoints are false minutiae and should be eliminated. Two endpoints are identified as a broken ridge structure by the following decision rules.(Fig.3(a)) (1) p (x1 − x2 )2 + (y1 − y2 )2 < Dist1 (1) (2) The line constructed by connecting two endpoints and two ridges connected with each minutia should all flow in the same direction tan−1 (
y2 − y1 1 ) ' (ORA + ORB ), x2 − x1 2
(2)
(3) Two ridges should be flowing to the opposite direction without being connected. For example(Fig.3(a)), if ridge connected with minutia A flows downwards,the other ridge should flow upwards and minutia B should be placed above the minutia A. 3.2
Detecting Bridge Structure
Due to excessive finger pressure or noise in the image, two separate ridges are sometimes connected by a short ridge to make a bridge structure. Based on the fact that ridges in fingerprint flow smooth and neighbor ridges flow in similar direction, method for detecting two false bifurcation associated with bridge structures is as follows. (Fig.3(b))
(1) Start tracking three ridges connected to a bifurcation(Point A). (2) If one of the tracked ridges meet another bifurcation(Point B), calculate orientation of the ridge connected by two bifurcations(ORAB ) and the distance between two bifurcations(DistAB ). (3) If the DistAB is less than a threshold value(Dist2 ) and the difference between the ORAB and the average orientation of two bifurcations(ORA , ORB ) is larger than a specified angle( π4 used in this dissertation),then two bifurcations are identified as a bridge structure. Note that by applying the described rule, false minutiae in triangular structures (Fig.2(g)) can also be detected efficiently. Two false minutiae and a true minutia are form a triangular structure. In the triangular structure, it is important to eliminate the two false minutiae while not eliminating the true minutia. By using the rule above, only two false minutiae are detected. 3.3
Detecting Short Ridge Structure
All short ridges should be considered as false minutiae because they are usually artifacts introduced by image preprocessing procedure such as ridge segmentation and thinning. To detect this kind of false minutiae, we start tracking ridges from ridge endings. If a tracked ridge meets another endpoint or a bifurcation within a distance(Dist3 ), two minutiae are considered as false minutiae. Also if a bifurcation meets another bifurcation while traking ridges and two bifurcations flow in opposite direction(Fig.2(e)), two bifurcations are considered false minutiae. 3.4
Detecting Hole Structure
Hole structures occur due to pores and dirts on fingerprints. The hole structure can be detected by tracking three ridges connected to an extracted bifurcation. If two tracked ridges meet to form another bifurcation and two bifurcations are within a distance(Dist4 ), then both bifurcations are considered as false minutiae. 3.5
Thresholds and False Minutiae Elimination
Because fingers are elasitic, distances between ridges change everytime due to different pressure a user puts on a input device. To cope with this problem, various thresholds used in this paper are made adaptive based on ridge frequency(Table.1). Ridge frequency is already calculated in image enhancement process[6]. To efficiently eliminate false minutiae while retaining true minutiae, false minutiae are detected and eliminated in specific order as shown in Fig.4.
Fig. 4. False Minutiae Elimination Order
Dist Dist1 Dist2 Dist3 Dist4
Description Broken Ridge Bridge Short Ridge Hole
Threshold 2/freq 1.5/freq 1.7/freq 2/freq
Table 1. Threshold Values : Freq indicates the ridge frequency at the minutia
DMR(%) EMR(%) TMR(%) FMR(%)
Method A 9.8 6.1 84.1 54.2
Method B 12.3 5.8 81.9 21.2
Table 2. Postprocessing performance: Method(A)- Raw Minutiae Extraction, Method(B)-Postprocessing adopted (DMR : Dropped Minutiae Ratio, EMR : Exchanged Minutiae Ratio, TMR : True Minutiae Ratio, FMR : False Minutiae Ratio)
4
Experimental Results
In this section, the performance of the proposed minutiae postprocessing algorithm is evaluated. Fingerprint images were acquired through optic-based fingerprint sensor manufactured by Nitgen. The size of the image is 248 × 292 with the resolution of 450 dpi and 1000 fingerprint images(10 fingerprints for 100 individuals) with various image qualities were used for experiments. Before showing the experimental results, we will describe some terms used to evaluate the performance. True Minutiae(TM) are minutiae picked by an expert. Paired Minutiae(PM) are minutiae extracted by the system which coincide with TM. False Minutiae(FM) are minutiae extracted by the system which do not coincide with TM. Dropped Minutiae(DM) are minutiae picked by an expert which are not extracted by the system. Finally, Exchanged Minutiae(EM) are minutiae extracted by the system which coincide with TM except the type. Table 2 shows the performance of our proposed minutiae postprocessing algorithm. The method A indicates the results of the raw minutiae extraction(without postprocessing) and the method B indicates the results when our postprocessing algorithm is adopted. It shows that the False Minutiae Ratio drops 33% while the True Minutiae Ratio only drops 2.2%. In addition, to see the effect of our postprocessing algorithm on actual fingerprint matching, we adopted matching procedure from [8]. Fig.5 shows matching results with ROC curves. It is clear from the results that the performance of the matching system is greatly improved by adopting the proposed postprocessing algorithm.
Fig. 5. ROC(Receiver Operating Characteristic) Curves
5
Conclusion
A minutiae postprocessing algorithm was proposed in this paper. To avoid eliminating true minutiae while postprocessing, our proposed algorithm was based on the orientation and flow of ridges as well as minutiae distance and connectivity. Experimental results showed that our algorithm is indeed very effective; eliminating great deal of false minutiae while retaining most of true minutiae. It was also shown that the proposed algorithm improves the fingerprint matching performance.
References [1] D. Maio, and D. Maltoni, “Direct Gray-Scale Minutiae Detection in Fingerprints,“ IEEE Trans. Pattern Anal. Machine Intell., vol. 19, no. 1, pp. 27 - 39, 1997 [2] A. K. Jain, L. Hong, and R. Bolle, “On-Line Fingerprint Verification,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 19, no. 4, pp.302-313, April 1997. [3] L. C. Jain, et al, Eds., Intelligent Biometric Techniques in Fingerprint and Face Recognition, CRC Press International Series on Computational Intelligence, 1999. [4] N. K. Ratha, S. Chen, and A. K. Jain, “Adaptive Flow Orientation Based Feature Extraction in Fingerprint Images,“ Pattern Recognition, vol. 28, no. 11, pp. 1,657 1,672, 1995 [5] Q. Xiao, and H. Raafat, “Fingerprint Image Postprocessing : A Combined Statistical and Structural Approach,“ Pattern Recognition, vol. 28, no. 11, pp. 1,657 1,672, 1995 [6] L. Hong, Y. Wan, and A. K. Jain, “Fingerprint Image Enhancement: Algorithm and Performance Evaluation,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol.20, pp.777–789, Aug. 1998. [7] J. R. Parker, Algorithms for Image Processing and Computer Vision, New York : Wiley Computer Publishing, 1997. [8] A. Wahab, S. H. Chin, E. C. Tain, “Novel Approach to Automated Fingerprint Recognition,” IEE Proc.- Vis. Image Signal Process, vol.145, no.3, pp.160–166, Jun. 1998.
Abstract. A common problem in fingerprint recognition is the existence of false minutiae which increase both FAR and FRR in fingerprint matching. In this paper, a robust minutiae postprocessing algorithm is proposed. Unlike most algorithms which use simple distance and connectivity criterions for postprocessing, we also used orientation and flow of ridges as the key factor for postprocessing to avoid eliminating true minutiae while postprocessing. It is shown by the experiments that our postprocessing algorithm improves the minutiae extraction accuracy and the performance of the matching process.
1
Introduction
Most fingerprint recognition systems are based on minutiae matching [1]. Minutiae are local discontinuities of fingerprints and are restricted to two types : ridge ending and ridge bifurcation [2], [3]. A common problem in fingerprint recognition is the existence of false minutiae which increase both FAR and FRR in fingerprint matching. Therefore, the enhancement of the fingerprint image and the false minutiae elimination form an important part of the system. However, most of the researches emphasized on the fingerprint image enhancement and the false minutiae elimination process was based on simple distance and connectivity criteria [1], [2], [3], [4]. But the problem with such simple approaches is that it eliminates true minutiae while eliminating false minutiae. Xiao and Raafat proposed in [5], a minutiae postprocessing algorithm based on both statistical and structural information. However, their method relies heavily on connectivity which makes it complex and unreliable to bad quality fingerprints. Also, specific structural informations were not given. In this paper, we present an efficient minutiae postprocessing algorithm. The goal of our minutiae extraction algorithm is to remove as many false minutiae as possible while retaining true minutiae. The goal is achieved by postprocessing minutiae based on not only the minutiae distance and connectivity but also using the orientation and flow of ridges as the key factor. Rest of the paper is organized as follows. Section 2 briefly decribes the adopted preprocessing procedures. Section 3 decribes the proposed minutiae postprocessing procedures. The performance of the proposed algorithm is shown by experiments in Section 4. Finally, Section 5 contains conclusion.
2 2.1
Preprocessing and Minutiae Extraction Preprocessing
Preprocessing procedures necessary for minutiae extraction are shown in Fig.1.
Fig. 1. Preprocessing Procedures
The first preprocessing procedure is the calculation of the local ridge orientation. The least mean square orientation estimation algorithm[6] is used and the local ridge orientation is specified by blocks rather than every pixel. The calculated orientation is in the range between 0 and π After the ridge orientation calculation, ridge frequency is calculated[6]. Using the calculated orientations and frequencies, the input grayscale image is enhanced and binarized by Gabor filters which have both frequency-selective and orientation-selective properties[6]. The final preprocessing operation required before extracting minutiae is thinning. Thinning reduces the the widths of the binary ridges down to a single pixel to facilitate the job of detecting ridge endings and bifurcations. The Zhang-Seun thinning algorithm[7] is used in this paper. 2.2
Minutiae Extraction
After a thinned fingerprint image is obtained, minutiae are directly extracted from the thinned image. To detect minutiae, a count of the pixel value transition at a point of interest in a 3×3 mask is used [4], [5]. If the count equals 2, then the point is an endpoint. If the count equals 6, then the point is a bifurcation. For each extracted minutia, the x & y coordinate and the orientation are recorded. The minutiae orientation is defined as the local ridge orientation of the associated ridge[2]. The minutiae orientation is in the range between 0 and π.
3
Proposed Minutiae Postprocessing Algorithm
There are many false minutiae among the extracted minutiae. False minutiae will decrease the performance of the fingerprint identification system by increasing both FRR and FAR. Typical false minutiae structures are shown in Fig.2. In this section, minutiae postprocessing algorithm is proposed. To eliminate false minutiae without eliminating true minutiae, the proposed methods are based on the flow of ridges as well as the minutiae distance and connectivity.
(a)
(b)
(c)
(d)
(e)
(f)
(g)
Fig. 2. Typical false minutiae : (a)Broken ridge, (b)Bridge, (c)Short ridge, (d)Short ridge, (e) Short Ridge, (f) Hole, (g)Triangle
(a)
(b)
Fig. 3. False Minutiae Structures: (a)Broken ridge, (b)Bridge
3.1
Detecting Broken Ridge Structure
Because of scars and insufficient finger pressure on the input device, a ridge may break into two ridges creating two endpoints. Obviously, these two endpoints are false minutiae and should be eliminated. Two endpoints are identified as a broken ridge structure by the following decision rules.(Fig.3(a)) (1) p (x1 − x2 )2 + (y1 − y2 )2 < Dist1 (1) (2) The line constructed by connecting two endpoints and two ridges connected with each minutia should all flow in the same direction tan−1 (
y2 − y1 1 ) ' (ORA + ORB ), x2 − x1 2
(2)
(3) Two ridges should be flowing to the opposite direction without being connected. For example(Fig.3(a)), if ridge connected with minutia A flows downwards,the other ridge should flow upwards and minutia B should be placed above the minutia A. 3.2
Detecting Bridge Structure
Due to excessive finger pressure or noise in the image, two separate ridges are sometimes connected by a short ridge to make a bridge structure. Based on the fact that ridges in fingerprint flow smooth and neighbor ridges flow in similar direction, method for detecting two false bifurcation associated with bridge structures is as follows. (Fig.3(b))
(1) Start tracking three ridges connected to a bifurcation(Point A). (2) If one of the tracked ridges meet another bifurcation(Point B), calculate orientation of the ridge connected by two bifurcations(ORAB ) and the distance between two bifurcations(DistAB ). (3) If the DistAB is less than a threshold value(Dist2 ) and the difference between the ORAB and the average orientation of two bifurcations(ORA , ORB ) is larger than a specified angle( π4 used in this dissertation),then two bifurcations are identified as a bridge structure. Note that by applying the described rule, false minutiae in triangular structures (Fig.2(g)) can also be detected efficiently. Two false minutiae and a true minutia are form a triangular structure. In the triangular structure, it is important to eliminate the two false minutiae while not eliminating the true minutia. By using the rule above, only two false minutiae are detected. 3.3
Detecting Short Ridge Structure
All short ridges should be considered as false minutiae because they are usually artifacts introduced by image preprocessing procedure such as ridge segmentation and thinning. To detect this kind of false minutiae, we start tracking ridges from ridge endings. If a tracked ridge meets another endpoint or a bifurcation within a distance(Dist3 ), two minutiae are considered as false minutiae. Also if a bifurcation meets another bifurcation while traking ridges and two bifurcations flow in opposite direction(Fig.2(e)), two bifurcations are considered false minutiae. 3.4
Detecting Hole Structure
Hole structures occur due to pores and dirts on fingerprints. The hole structure can be detected by tracking three ridges connected to an extracted bifurcation. If two tracked ridges meet to form another bifurcation and two bifurcations are within a distance(Dist4 ), then both bifurcations are considered as false minutiae. 3.5
Thresholds and False Minutiae Elimination
Because fingers are elasitic, distances between ridges change everytime due to different pressure a user puts on a input device. To cope with this problem, various thresholds used in this paper are made adaptive based on ridge frequency(Table.1). Ridge frequency is already calculated in image enhancement process[6]. To efficiently eliminate false minutiae while retaining true minutiae, false minutiae are detected and eliminated in specific order as shown in Fig.4.
Fig. 4. False Minutiae Elimination Order
Dist Dist1 Dist2 Dist3 Dist4
Description Broken Ridge Bridge Short Ridge Hole
Threshold 2/freq 1.5/freq 1.7/freq 2/freq
Table 1. Threshold Values : Freq indicates the ridge frequency at the minutia
DMR(%) EMR(%) TMR(%) FMR(%)
Method A 9.8 6.1 84.1 54.2
Method B 12.3 5.8 81.9 21.2
Table 2. Postprocessing performance: Method(A)- Raw Minutiae Extraction, Method(B)-Postprocessing adopted (DMR : Dropped Minutiae Ratio, EMR : Exchanged Minutiae Ratio, TMR : True Minutiae Ratio, FMR : False Minutiae Ratio)
4
Experimental Results
In this section, the performance of the proposed minutiae postprocessing algorithm is evaluated. Fingerprint images were acquired through optic-based fingerprint sensor manufactured by Nitgen. The size of the image is 248 × 292 with the resolution of 450 dpi and 1000 fingerprint images(10 fingerprints for 100 individuals) with various image qualities were used for experiments. Before showing the experimental results, we will describe some terms used to evaluate the performance. True Minutiae(TM) are minutiae picked by an expert. Paired Minutiae(PM) are minutiae extracted by the system which coincide with TM. False Minutiae(FM) are minutiae extracted by the system which do not coincide with TM. Dropped Minutiae(DM) are minutiae picked by an expert which are not extracted by the system. Finally, Exchanged Minutiae(EM) are minutiae extracted by the system which coincide with TM except the type. Table 2 shows the performance of our proposed minutiae postprocessing algorithm. The method A indicates the results of the raw minutiae extraction(without postprocessing) and the method B indicates the results when our postprocessing algorithm is adopted. It shows that the False Minutiae Ratio drops 33% while the True Minutiae Ratio only drops 2.2%. In addition, to see the effect of our postprocessing algorithm on actual fingerprint matching, we adopted matching procedure from [8]. Fig.5 shows matching results with ROC curves. It is clear from the results that the performance of the matching system is greatly improved by adopting the proposed postprocessing algorithm.
Fig. 5. ROC(Receiver Operating Characteristic) Curves
5
Conclusion
A minutiae postprocessing algorithm was proposed in this paper. To avoid eliminating true minutiae while postprocessing, our proposed algorithm was based on the orientation and flow of ridges as well as minutiae distance and connectivity. Experimental results showed that our algorithm is indeed very effective; eliminating great deal of false minutiae while retaining most of true minutiae. It was also shown that the proposed algorithm improves the fingerprint matching performance.
References [1] D. Maio, and D. Maltoni, “Direct Gray-Scale Minutiae Detection in Fingerprints,“ IEEE Trans. Pattern Anal. Machine Intell., vol. 19, no. 1, pp. 27 - 39, 1997 [2] A. K. Jain, L. Hong, and R. Bolle, “On-Line Fingerprint Verification,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 19, no. 4, pp.302-313, April 1997. [3] L. C. Jain, et al, Eds., Intelligent Biometric Techniques in Fingerprint and Face Recognition, CRC Press International Series on Computational Intelligence, 1999. [4] N. K. Ratha, S. Chen, and A. K. Jain, “Adaptive Flow Orientation Based Feature Extraction in Fingerprint Images,“ Pattern Recognition, vol. 28, no. 11, pp. 1,657 1,672, 1995 [5] Q. Xiao, and H. Raafat, “Fingerprint Image Postprocessing : A Combined Statistical and Structural Approach,“ Pattern Recognition, vol. 28, no. 11, pp. 1,657 1,672, 1995 [6] L. Hong, Y. Wan, and A. K. Jain, “Fingerprint Image Enhancement: Algorithm and Performance Evaluation,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol.20, pp.777–789, Aug. 1998. [7] J. R. Parker, Algorithms for Image Processing and Computer Vision, New York : Wiley Computer Publishing, 1997. [8] A. Wahab, S. H. Chin, E. C. Tain, “Novel Approach to Automated Fingerprint Recognition,” IEE Proc.- Vis. Image Signal Process, vol.145, no.3, pp.160–166, Jun. 1998.
