Biometrics: Fingerprint Classification
Team Manager : Jin-Hyuk Hong (hjinh_at_cs.yonsei.ac.kr)
Team Members: Jun-Ki Min (loomlike_at_sclab.yonsei.ac.kr), Ung-Keun Cho (bearoot_at_sclab.yonsei.ac.kr)
Fingerprint classification is an important stage in automatic fingerprint identification systems (AFIS) to reduce the number of one-to-one comparisons performed when executing a fingerprint query by categorizing it. According to the Henry system, fingerprints are often partitioned into several classes such as whorl, left loop, right loop, arch and tented arch.
Since the Henry system categorizes fingerprints by the relative position and number of core and delta points in the print, many researchers have tried to extract the singular points in the flow of the ridges. Karu and Jain proposed a heuristic algorithm with singularities, Nyongesa, et al. used the relative positions of cores and deltas while Zhang and Yan used singularities together with pseudo ridges to classify finger-prints. Even though singularities can provide brief classification based on rules designed by experts, it is hard to obtain high accuracy because of the quality of fin-gerprint images.
In order to obtain a high classification rate, various features, such as FingerCode, ridge distributions and directional images, have been also actively investigated. Jain, et al. proposed FingerCode that uses a Gabor filter to extract the directional ridge flow, and Park used the orientation filtered by fast Fourier transform. Cappelli, et al. proposed the directional image which models a fingerprint with a graph. Nagaty extracted a string of symbols using the block directional image of a fingerprint, while Chang and Fan proposed the ridge distribution model which consists of a com-bination of 10 basic ridge patterns with different ridge distribution sequences. Since these features include more information than simple singularities, fingerprint classification has become more accurate.
There are some other attempts to integrate several features and methods to produce a robust fingerprint classifier. Senior used hidden Markov models and decision trees to recognize the ridge structure of the print, while Yao, et al. combined flat and structured features using the recursive neural networks and support vector ma-chines (SVMs). By combining several types of features and methods, fingerprint classification might be more accurate and reliable when there are some noises in-cluded.
- A probabilistic classification method based on the analysis of both singularities and pseudo ridges using a naive Bayes classifier: Singularities and pseudo ridges are used for fingerprint classification, but it is likely to classify into a wrong class by the rule-based classification due to features misextracted. Since appropriate model for these structural features is required, the naive Bayes classifier is used for it in this method. It classifies samples probabilistically by using the features as evidences of the inference, while the naive Bayes classifier imposes a strong assumption of independence among features.
- Multiple decision templates (MuDTs): The method extracts several clusters of different characteristics from each class of fingerprints and constructs localized models to overcome the restriction of the single model, which may be weak at ambiguous samples. The outputs of one-vs-all SVMs for the training data are clustered by the SOM to decompose the class into several clusters to separate diverse characteristics. The localized decision templates are estimated for each cluster, and then the MuDTs are constructed.
- Dynamic fingerprint classification: This method is a novel fingerprint classification approach integrating the naive Bayes classifier (NB) and SVMs that use different fingerprint features. In order to accomplish highly accurate classification, SVMs with FingerCode are generated based on the one-vs-all (OVA) scheme while NB with singularities dynamically organizes them. Since some fingerprints are hard to discriminate their class, OVA SVMs are sequentially evaluated ordered by the probability of classes from NB to avoid the ambiguity.
3. Experimental Results
The NIST Special Database 4 was used to verify the proposed method. The NIST DB 4 consists of 4 000 fingerprint images (2 000 pairs), which are equally distributed into 5 classes (W, R, L, A, T). Some ambiguous images are hard to classify because of scars in the fingerprints, low quality images, or ridge structures that are characteristic of two different classes. The NIST DB 4 contains 350 fingerprint pairs that can be assigned to more than one class. We use the first label in training, and considered the second label together in our tests.
 E.-K. Yun, J.-H. Hong, and S.-B. Cho, "Adaptive enhancing of fingerprint image with image characteristics analysis," Lecture Notes in Artificial Intelligence (Proc. of the 17th Australian Conf. on Artificial Intelligence), vol. 3339, pp. 120-131, 2004.
 J.-H. Hong, E.-K. Yun and S.-B. Cho, "A review of performance evaluation for biometrics systems,"
Int. J. of Image and Graphics, vol. 5, no. 2, pp. 501-536, 2005.
 J.-K. Min, J.-H. Hong and S.-B. Cho, "Effective fingerprint classification by localized models of support vector machines," Lecture Notes in Computer Science (ICBA), 2005.
 J.-K. Min, J.-H. Hong and S.-B. Cho, "SVM의 다중결정템플릿을 이용한 지문분류," J. Korea Information Science Society, 2005.