Abstract
Fingerprint recognition is a time consuming process in huge fingerprint data-bases. Classifying fingerprints and narrowing the search to a specific class is an effective way to reduce the processing time. For the classification of noisy fingerprints acquired by live scan devices, we proposed a new scheme, based on statistical features derived from dense scale invariant feature transform (d-SIFT). Dense SIFT descriptor is not computed for certain key points but for all pixels of a fingerprint. Each SIFT descriptor consists of 128 features and the distribution of each feature over all SIFT descriptors extracted from a finger-print is used to compute statistical features (mean, standard deviation, kurtosis and skewness). In this approach the dimension of the feature space remains fixed (4x128) irrespective of the number of SIFT descriptors. Extreme Learning Machine (ELM) with radial basis function (RBF) kernel is used as a classifier. The proposed scheme has been evaluated using FVC2004, a bench-mark database of noisy low quality live scanned fingerprints; it achieved the average accuracy of 97.63, which is better than the state-of-the-art fingerprint classification methods.