Minutia Verication and Classication for Fingerprint Matching

Salil Prabhakar, Anil K. Jain, Jianguo Wang Sharath Pankanti & Ruud Bolle Dept. of Comp. Sci. and Eng. IBM T. J. Watson Research Center Michigan State University, Yorktown Heights, NY 10598 East Lansing, MI 48824 prabhaka,jain,wangjia3 @cse.msu.edu sharat,bolle @watson.ibm.com

Abstract
Raw image data offer rich source of information for matching and classication. For simplicity of pattern recognition system design, a sequential approach consisting of sensing, feature extraction and matching is conventionally adopted where each stage transforms a particular component of information relatively independently. The interaction between these modules is limited. Some of the errors in the end-to-end sequential processing can be easily eliminated especially for the feature extraction stage by revisiting the original image data. We propose a feedback path for the feature extraction stage, followed by a feature renement stage for improving the matching performance. This performance improvement is illustrated in the context of a minutiae-based ngerprint verication system. We show that a minutia verication stage based on reexamining the gray-scale prole in a detected minutias spatial neighborhood in the sensed image can improve the matching perforon our database. Further, we show that a mance by feature renement stage which assigns a class label to each detected minutia (ridge ending and ridge bifurcation) before matching can also improve the matching performance . A combination of feedback (minutia vericaby tion) in the feature extraction phase and feature renement (minutia classication) improves the overall performance of . the ngerprint verication system by

each component (stage) is not sufcient/effective to capture the essence of the sensed data. One of the problems with the staged approach is that the limited use of information in each stage results in feature extraction and/or matching performance artifacts. Thus, though the sequential (staged) approach is efcient from design and processing point of view, it may introduce errors in the feature extraction and recognition results. We believe that by revisiting the original image data, some of the mistakes in the end-to-end sequential processing can be eliminated, resulting in an improvement in system performance. Performance can also be improved by feature renement. See Figure 1 for our proposed changes to a sequential feature extraction system. We illustrate the above approach in the ngerprint matching domain. Most of the existing automatic nger-

Feedback

Sensed image

Feature Extraction

1 Introduction
The human visual system relies on image data for decision making because of the richness of the image context. Ideally, we would like to design pattern recognition systems which make decisions based on all the information available in the input image. However, traditionally, for simplicity of design, a sequential approach is often adopted to feature extraction and matching where each stage transforms a particular component of the information relatively independently and the interaction between these components of information is limited. Often, the rather simplistic model used in 1

Matching

Feature Refinement

Figure 1. A general pattern recognition system with proposed feedback in feature extraction and a new feature renement stage. print verication systems rst detect the minutiae in a ngerprint image and then match the input minutiae set with the stored template [1, 2]. A typical algorithm described in [1] uses a sequential approach to feature extraction (see

Figure 2). The feature extraction rst binarizes the ridges in a ngerprint image using masks that are capable of adaptively accentuating the local maximum gray-level values along a direction normal to the local ridge direction. Minutiae (ridge bifurcation and ending; see Figure 3) are determined as points that have one neighbor or more than two neighbors in the skeletonized image. However, the orientation estimation in a poor quality image is extremely unreliable, resulting in the detection of many false minutiae. Several researchers have proposed minutia-pruning in the postprocessing stage to delete spurious minutiae [1, 3, 4] but the pruning is based on rather ad-hoc techniques. In this paOriginal Image Orientation field Binarization

in the feedback system (minutia verication) is to learn the characteristics of minutiae in gray level images which can then be used to verify each detected minutia. This step will replace the rather ad-hoc minutia-pruning stage used in [1]. Each detected minutia is ltered through this verication stage and is either accepted or rejected based on the learnt gray level characteristics in the neighborhood of a minutia. The minutia classier is based on supervised training using Learning Vector Quantization [5]. We show that the feature renement (minutia classication into bifurcation and ending) can further improve the matching performance. We use a rule-based classier to classify a minutia into the two categories. The matching algorithm proposed in [1] is modied to match minutiae of the same type in the sensed image and the template. The modication of minutia matching algorithm used in [1] with minutia verication and minutia classication signicantly improves the matching accuracy.

2 Minutia Verication
Our minutia verication algorithm can be divided into feature extraction, training (learning three stages; verication. the minutiae characteristics), and

2.1 Feature Extraction

Figure 2. Minutiae extraction algorithm.

Figure 3. Examples of ngerprint minutiae; ridge endings () and bifurcations (). per, we propose a feedback system for minutiae extraction which is based on an analysis of the gray scale prole in the neighborhood of potential minutiae. We also propose a feature renement stage where the minutiae are classied into two major classes: ridge bifurcation and ending. The goal 2

We use the minutiae detection algorithm developed by Jain et al. [1] for our study. Each detected minutia has the following three attributes: the and position and the direction of the ridge in which the minutia resides. We extract region centered at the and position of the a minutia and oriented in the direction of the minutia. The extracted region is normalized to a constant mean and variance to remove the effects of sensor noise and gray-scale deformation because of nger pressure variations. In our experiments, we set the values of both mean and variance . We enhance the contrast of the ridges by lterto window with a appropriately tuned Gaing each bor lter [6]. We set the frequency of the Gabor lter to the average ridge frequency ( ), where is the average inter-ridge distance. The average inter-ridge distance is appixels in a ngerprint image. The proximately values of parameters and for Gabor lters were em(about half the pirically determined and each is set to average inter-ridge distance). Since the extracted region is in the direction of the minutia, the lter is tuned to direction. We perform the ltering in the spatial domain with a mask size of . The Gabor response for each pixel in the region is scaled to eight gray levels. We extract a region (see Figure 4) from the center of the region to avoid boundary problems in normalization and l-dimensional tering and concatenate the rows to form a feature vector.

2.2 Training
In the training phase, minutiae and non-minutiae feature vectors are fed to a Learning Vector Quantizer to learn the characteristics of minutiae and non-minutiae regions. We nuse the IBM HURSLEY database that contains different ngers for training and gerprint images from testing. The multiple impressions for each nger in the database were taken at signicantly different times. The images are of different sizes but all the images have been resolution with gray levels. A nscanned at gerprint expert has marked the true ngerprint minutiae in images for training and these images. We use the rst images from different ngers for testing. the remaining feature vectors correWe extract approximately sponding to all the true minutiae from the images in the training database. We also extract an equal number of negative samples (non-minutiae) by randomly sampling the images in the training set and making sure that there is no minutia in its immediate neighborhood. For the true minutia, we use the direction of the minutia provided by the expert. For the negative examples, we compute the direction of the block using the hierarchical orientation-eld algorithm [1].

(a)

(b)

Figure 4. Examples of gray level proles in the neighborhood of (a) minutiae and (b) nonminutiae. These subimages, scaled to to 8 gray levels, are used for training a LVQ.

2.3 Testing
We use two methods to test the LVQ-based minutiae vs. non-minutiae classier. In the rst method, we evaluate the classier using the ground truth minutia information in the test database. In the second method, we extract the minutiae from the test database using the minutiae extraction algorithm described in [1]. An automatically detected minutia may be slightly perturbed from its original location because of the noise introduced during the binarizing and thinning windows in processes. So, we extract twenty ve the neighborhood of each detected minutia and verify each window. The decisions from the verication of these windows are combined in a simple manner. If the classier yields a positive verication for any of the windows, the minutia is accepted. Figures 5 (a)-(c) compare the minutiae detection without pruning, with pruning, and with pruning replaced with minutia verication for a good quality ngerprint. Minutia verication is more effective for poor quality ngerprints.

improve the overall matching performance. In minutiae extraction algorithm, if a pixel in the thinned image has only one neighbor then the minutia is classied as an ending, and if a pixel has more than 2 neighbors, then the minutia is classied as a bifurcation. The matching algorithm in [1] is modied to match endings only with endings and bifurcations only with bifurcations. In our experience, there are signicantly more number of bifurcations present in a typical ngerprint than endings. See Figure 5 (d) for minutia classication results.

4 Experimental Results
We rst evaluated the performance of a minutiae-based ngerprint verication system [1]. The test consisted of evaluating the LVQ-based classier on the ground truth -dimensional feature vecdata. Approximately tors each for minutiae and non-minutiae were extracted from the images in the training database. The testing was done on feature vectors extracted only from the minutiae samples in the test database. The trained LVQ-based classier with one hundred code book vectors per class gives an on the training data and on the accuracy of test data. A real test for the utility of the verication module is the gain in matching accuracy. So, we replaced the minutia-pruning stage in the algorithm in [1] with the proposed minutia verication stage. In the ROC curve shown in Figure 6, the dotted line represents the matching accuracy on the test set and the solid line represents the performance when the pruning stage in [1] is replaced with the proposed minutia verication scheme. These ROC curves show that the overall performance of the ngerprint verication sysat all the operating points. tem increases by The benets of using minutia type information is illustrated in Figure 7. The solid line in the gure represents the performance when the minutia type information is used. Figure 8 shows the performance improvement when minu-

3 Minutia Classication
The American National Standards Institute proposes four classes of minutia: ending, bifurcation, trifurcation, and undetermined. The most discriminable categories are ridge ending and bifurcation (see Figure 3). Most of the ngerprint matching algorithms do not use minutia type information because the two types of minutiae get interchanged due to nger pressure difference or noise. However, we show that a consistent classication of minutia can indeed 3

(a)

(b)

(c)

(d)

Figure 5. Minutiae detection and classication; (a) Minutiae detection using the algorithm in [1] without pruning, (b) results of minutia-pruning, minutiae marked with red were pruned, (c) result of minutia verication instead of pruning, minutiae marked with red were rejected, (d) result of classifying minutiae shown in (b); minutia bifurcations are marked with blue and endings are marked with red.

100 95

100 95

Genuine Acceptance Rate (%)

90 85 80 75 70 65 With minutia verification With minutia pruning 60 3 10 10

Genuine Acceptance Rate (%)

90 85 80 75 70 65 With minutia classification and verification Without minutiae classification and verification 60 3 10

10 False Acceptance Rate (%)

10

10

10

10 False Acceptance Rate (%)

10

10

Figure 6. ROC for ngerprint matching when minutia verication is used.

Figure 8. ROC for ngerprint verication when both minutia classication and verication are used. mation) before matching. The proposed minutia verication scheme can be extended for detection of minutiae in a ngerprint image. We are working on a continuous classication of the minutiae into several categories, one of the categories being non-minutiae. A classication label and a condence value will be assigned to each minutiae and we will modify the matching algorithm to account for the condence values.

tia verication and classication are combined. The classication is done before the verication but the classication information is not used during the verication. The performance of the ngerprint verication system in [1] is signicantly improved by using minutia classication and verication.
100 95

Genuine Acceptance Rate (%)

90 85 80 75 70 65 With minutia type information Without minutia type information 60 3 10 10

References
[1] A. K. Jain, L. Hong, S. Pankanti, and R. Bolle, An Identity Authentication System using Fingerprints, Proceedings of the IEEE, Vol. 85, No. 9, pp. 13651388, 1997. [2] D. Maio and D. Maltoni, Direct Gray-Scale Minutiae Detection in Fingerprints, IEEE Trans. Pattern Anal. Machine Intell., Vol. 19, No. 1, pp. 27-40, 1997. [3] D.C.D. Hung, Enhancement and Feature Purication of Fingerprint Images, Pattern Recognition, vol. 26, no. 11, pp. 1,661-1,671, 1993. [4] Q. Xiao and H. Raafat, Fingerprint Image Postprocessing: A Combined Statistical and Structural Approach, Pattern Recognition, vol. 24, no. 10, pp. 985992, 1991. [5] T. Kohonen, J. Kangas, J. Laaksonen, and K. Torkkola, LVQ PAK: A Program Package for the Correct Application of Learning Vector Quantization Algorithms, in Proc. Intl Joint Conf. on Neural Networks, (Baltimore), pp. 1725-1730, June 1992. [6] A. K. Jain, S. Prabhakar, L. Hong, and S. Pankanti, Filterbank-based Fingerprint Matching, to appear in IEEE Trans. Image Processing, 2000.

10 False Acceptance Rate (%)

10

10

Figure 7. ROC for ngerprint matching when minutia classication is used.

5 Discussions and future work