After the above methodology was used to compare high quality
and low quality (noisy) images, all low quality prints (see Fig. 2 for
example images) were decomposed into individual component
frequencies using the discrete Fourier transform (DFT). This
transformation assumes that a complex signal can be decomposed
into a summation or integral of sines and cosines with varying
frequency and amplitude. To illustrate, consider a 1D square-wave
or step-function (Fig. 3 adapted from [22]). With regard to 2D
spatial images, the same theory applies; a 2D image can be
decomposed into component frequencies, revealing a power
spectrum. The power spectrum highlights repetitive patterns such
as parallel lines or ink dots in the spatial image, which manifest as
bright spots or lines in the power spectrum [23]. If selected
frequencies in the power spectrum are suppressed (or filtered) the
newly modified spectrum can be transformed back to the spatial
domain, often reducing or removing background interference. The
minutiae in the post-processed fingerprint image can then be reexamined,
triangulated, registered and compared to the high
quality exemplars again, providing match scores that can be
compared to those obtained pre-filtering.
The results of all pairwise comparisons were used to generate
probability density functions describing the match scores and the
number of matching minutiae based on the type of pairwise
comparison (e.g., known-matches (KM) and known non-matches
(KNM)), and further divided by subjectively categorized quality
(i.e., high, pre-filtered low and post-filtered low). In addition to
inspection of the aforementioned probability density functions, a
direct measure of the normalized gain in similarity was computed
for sets of pre- and post-filtered images according to Eq. (2) [24