1D Spectral Image Validation/Verification Metric for Fingerprints

摘要

Image validation and verification are important functions in the acquisition of fingerprint images from live-scan devices and for assessing and maintaining the fidelity of fingerprint image databases. Such databases are used by law enforcement agencies, for which data integrity is paramount, and many hours must be devoted to visual inspection of images. In addition, such databases are used by the National Institute of Standards and Technology (NIST) and others to test automated fingerprint identification system (AFIS) algorithms and to aide the advance of this technology. We propose a comparatively simple computational mechanism by which to screen fingerprint image databases for specimens improperly scanned from fingerprint cards, guide the auto-capture process and flag auto-capture failures, identify non-fingerprint images that may have been included in a database, and recognize aberrant sampling of fingerprint images. The scheme reduces an input image to a 1-dimensional power spectrum that makes explicit the characteristic ridge structure of the fingerprint that on a global basis differentiates it from most other images. The magnitude of the distinctive spectral feature, related directly to the distinctness of the level 1 ridge flow, provides a primary diagnostic indicator of the presence of a fingerprint image. The frequency of the spectral feature provides a secondary classification metric and, on a coarse level, indicates the scan sample rate of the fingerprint image. Test results are reported in which the Spectral Image Validation and Verification (SIVV) utility is applied to a variety of databases composed of fingerprint and non-fingerprint images. Using only the peak height and frequency limit as simple classification criteria, the SIVV utility achieves an equal error rate (EER) for false positive and false negative classifications of 10% for fingerprints mixed with a variety of non-fingerprint images, including many chosen to exhibit periodic structure similar to that of a fingerprint. An EER of around 7 % is found with a dataset containing fingerprints mixed with other biometric samples, i.e. face and iris images.