A Fingerprint Indexing Approach Using Multiple Similarity Measures and Spectral Clustering

Fingerprint identification is still an open problem. Many proposed solutions relies on expert's predefined classes. Recent studies and experiments have shown that predefined classes are not suitable for large databases. The main challenge is the distribution of predefined expert's classes. Most fingerprints are loops; this means an identification system designed based on this classical approach slows down a matching process due to many 1-1 comparisons that need to be done to find an exact match. Beside this, some fingerprints sometimes lose their core points. In such case, it becomes very difficult to find the right class. To solve this problem, minutiae points positions (x,y) and orientations (θ) are proposed in this paper. Minutiae points are extracted and passed to a feature dimensions reduction algorithm composed of locality sensitive hashing (LSH) and histograms. The algorithm proposed creates 260 fixed length feature vectors for each fingerprint. Feature vectors are then passed to a spectral clustering which automatically creates 25 classes. Classes are then indexed through three similarity measures, namely, Euclidean, Cosine and Minkowski distance. The performance of the proposed approach is compared with two fingerprint identification systems which are based on Minutiae Cylinder-Code (MCC) and Minutiae Quadruplets (MQ). The proposed approach is 89.6% accurate while MCC and MQ approaches gets 89.4% and 79.56%, respectively, on a NIST 4 special database.