Integrating minutiae based fingerprint matching with local correlation methods
MetadataShow full item record
Fingerprint is the most common biometric modality in use today for various civilian and security applications of access control. Fingerprint verification is the process of matching a pair (test and enrolled) of finger-skin ridge impressions to determine if the impressions are from the same finger. The non-linear stretching of the finger-skin makes any two instances of impressions of the same finger (quite) different. Other challenges are due to the varying contrast levels (moist fingers tend to give smudged impressions and dry fingers tend to give impressions with broken ridge contours), the rotation of the finger, the pressure applied on the sensor, and the partial nature of most fingerprints captured by commercial sensors. Fingerprint matching algorithms reported in the literature are of three types based on: (i) minutiae (discontinuities in the ridge contour) (ii) texture, and (iii) image correlation. Minutiae based matching methods consider special points of fingerprint impressions representing ends and bifurcation points of the fingerprint ridge structure. In texture matching, spatial relationship and geometrical attributes of the fingerprint ridges are used. Correlation scores from the intensities of corresponding pixels are used in the third approach. Although minutia based algorithms usually provide the best performance, they have problems matching partial or low quality fingerprint images when only a few minutiae are successfully extracted. Texture and correlation based matching methods have advantages dealing with such images as they utilize low level features which are not accounted for by minutia templates. On the other hand, minutia based approaches are faster and can take into consideration the non-linear deformations of fingerprint impressions. We have combined the strengths of the minutia and correlation approaches, into a single fingerprint matching algorithm. The first stage of our methodology consists of developing fast techniques to find correspondences between minutia sets utilizing fingerprint minutia indexing, and the second stage consists of verifying the results using correlation methods. We have investigated the effectiveness of integrating image correlation information between local neighborhood areas of matched minutiae, edges connecting the neighboring matched minutiae pairs, and the polygonal areas enclosed by neighboring minutia. We have used the Mutual Information as a similarity measure at the local level to improve the fingerprint matching rate because when the corresponding fingerprint images are aligned then the amount of information they contain about each other is maximal. We have also modeled the non-linear distortion caused by the stretching of the finger-skin and explored fusion techniques to combine both the global minutiae distribution structure and the local matching similarity. The novelty of our approach lies in the creation of a single non-linear correlation based matching algorithm. We have tested all the proposed methods on standard fingerprint benchmark datasets FVC2002 and reported the results superior to prior work.