The National Institute of Standards and Technology (NIST), with the cooperation of eight technology providers, performed a test of accuracy for searching latent fingerprints when using Automatic Feature Extraction and Matching (AFEM). This test is Phase II of the Evaluation of Latent Fingerprint Technology (ELFT) project. The test was open to both the commercial and academic community, and participants included vendors of Automated Fingerprint Identification Systems (AFIS). This report provides the design, process, caveats, results, observations and conclusions of the test.

The eight technology providers each submitted a Software Development Kit (SDK) containing a latent fingerprint and ten-print minutiae extraction algorithm, and a 1:n match algorithm that returns a candidate list report. The specific fingerprint features extracted by the SDK were at the discretion of the technology provider and could be proprietary, and the feature template input to the SDK’s matcher may include the original latent fingerprint image in its entirety. Technology providers were encouraged to submit research algorithms in this study. There was no requirement for the SDK’s to be in operational use or commercially available. NIST performed a pre-test of the SDK’s to ensure all functional capabilities were working. After validation of the SDK’s, the technology providers were no longer involved in the testing. NIST performed the same test on all SDK’s.

The test dataset contained 835 latent fingerprints, the associated ten-print fingerprint records containing the mates to the latent fingerprints, and two separate galleries of ten-print fingerprint records: one containing 5,000 records (50,000 fingerprints), and the second containing 10,000 records (100,000 fingerprints). The latent fingerprints were studied at two image resolutions: 1000 pixels per inch (ppi) (39.37 pixels per millimeter (ppmm)) images, and sub-sampled 1000 ppi producing 500 ppi (19.69 ppmm) images. In all tests, the ten-print galleries were 500 ppi. The technology providers had no knowledge of, or access to, the fingerprint datasets prior to, during, or after the tests.

In addition to assessing the overall performance of AFEM latent fingerprint technology, tests were designed to study specific factors expected to significantly impact performance. Insights into the effect of some of these factors may contribute to automated determination of latent fingerprint image quality. To this end, factors analyzed included the effect of gallery size, latent image resolution, supplementary region of interest, latent minutiae count, finger position, and finger pattern classification.
NIST performed analyses of the data and determined the performance and accuracy for each technology provider’s SDK. A summary of identification rates based on candidate list position (rank) is reported in the following table. Note that each latent fingerprint search generated a list of fifty candidates, and it was generally observed that most identifications occurred within the top ten. Therefore, rank one and rank top ten results are reported.