Fingerprint Image Compression Interoperability Analysis

In this project, we investigate the degree of interoperability among various fingerprint-matching algorithms using compressed enrollment (template) images compatible with ISO/IEC FCD 19794-3 [1]. The interoperability testing consists of three different fingerprint-matching algorithms: a pattern-based matching algorithm from Bioscrypt, and two minutiae-based matching algorithms from Innovatrics and Neurotechnologija. The 19794-3 compatible enrollment images are compressed with different parameter settings to assess the effect of these parameters. 1. FINGERPRINT DATABASE A database of images from a Crossmatch V300 sensor, collected during Q1 and Q2 of 2004 is used for all interoperability testing. The size of each image is 640× 480 (width × height) pixels. The resolution of each image is 500 dpi (dots per inch). The database consists of 150 fingers (LR, LM, LI, RI, RM, RR from each of 25 people). Each finger has 1 enrollment print and 10 verification prints. 2. COMPRESSION OF ENROLLMENT IMAGES The enrollment images are compressed using the pre-defined spectral-triple representation described in ISO/IEC FCD 19794-3. Below, we present a list of the particulars in our implementation. • ISO/IEC FCD 19794-3 section 5.2 particular: The image is downsampled from 500dpi to 200dpi using bilinear downsampling, and image cropping is not performed. Note that image cropping would reduce the data storage requirements and minimize the extraneous artifacts seen in the outer portions of figure 2(b). • ISO/IEC FCD 19794-3 section 5.3 particular: The image is partitioned into non-overlapping cells. • ISO/IEC FCD 19794-3 section 5.4 particular: The distance measure used to select the best fitting spectral-triple parameters is L1 norm (absolute distance). 3. TESTING PROCEDURES The testing procedures are summarized below: • Preprocess each enrollment image with an open source fingerprint image enhancement algorithm [2], [3] for ridge enhancement (an example is shown in figure 1) • Compress the enrollment image using the procedures described in section 2 • Store the compressed image • Reconstruct the enrollment image from its spectral-triple representation, and upsample the reconstructed image from 200dpi to 500dpi using bilinear interpolation (an example is shown in figure 2) • Match the reconstructed enrollment image against a verification print or an attacker print using each matching algorithm Attacker prints are considered to be the first print in the verification set for every finger other than the finger from which the current template is derived. This means that there are 149 attacker prints for every template. Since there are 10 verification prints per finger, each enrollment print set contains 1500 (150× 10) legitimate matching scores and 22350 (150 × 149) attacker matching scores.