generate a high normalized match score. Conversely, if the
questioned image is of poor quality, scores for known-matches
and known non-matches will be similar, and in the worst case
scenario, the KNM density will be widely distributed generating a
low normalized match score. Despite the theoretical simplicity in
the metric, the challenge is to (i) extract relevant details from the
questioned print and (ii) use these variables to predict the
magnitude of the match score without requiring computation of
the match score itself. To do this, the NFIQ determines image
attributes such as contrast, ridge flow, ridge curvature and
neighborhood statistics as a function of pixel intensity. Armed
with this information, extracted minutiae are weighted based on
the assessed quality of the local pixel block in which they belong.
The end result is an 11-dimensional feature vector that, through
the use of an artificial neural network, can be classified as poor (5),
fair (4), good (3), very good (2), or excellent (1) in quality [7].
Although extremely easy to use, the authors had concerns about
interpreting the expected results from the NIST metric prior to
initiating this phase of the research. The reason for this concern is
that the NFIQ was never designed to assess the types of images
examined in this study, especially given the extreme background
features purposely selected for inclusion in this project. This is not
a misgiving of the metric itself, but rather the nature of the
question being addressed by this study. The end result is that
caution must be exercised when interpreting the NFIQ results for
the images utilized in this project.