Master Thesis - Department of Computer Science

features (gradient coherence, intensity mean, and intensity variance) are computed
for each pixel, and a linear classifier associates the pixel with the background or
the foreground. A supervised technique is used to learn the optimal parameters
for the linear classifier for each specific acquisition sensor. A final morphological
post-processing step [43] is performed to eliminate holes in both the foreground and
background and to regularize the external silhouette of the fingerprint area. Their
experimental results showed that this method provides accurate results; however, it’s
computational complexity is markedly higher that most of the previously described
block-wise approaches.
2.2.1.2 Image Enhancement
The goal of an enhancement algorithm is to improve the clarity of the ridge structures
in the recoverable regions and mark the unrecoverable regions as too noisy for further
processing.
The most widely used technique technique for fingerprint image enhancement is
based on contextual filters. In contextual filtering, the filter characteristics change
according to the local context. In fingerprint enhancement, the context is often
defined by the local ridge orientation and local ridge frequency. The methods proposed
by O’ Gorman and Nickerson [90, 91] was one of the first to use contextual filtering.
They defined a mother filter based on four main parameters of fingerprint images at a
given resolution; minimum and maximum ridge width, and minimum and maximum
valley width. The local ridge frequency is assumed constant and therefore, the context
is defined only by the local ridge orientation.
Sherlock, Monro, and Millard [110, 111] performed contextual filtering in the
Fourier domain. The filter defined in the frequency domain is the function:
H(ρ, θ) = Hradial(ρ).Hangle(θ), (2.19)
where Hradial depends only on the local ridge spacing ρ = 1/f and Hangle depends
only on the local ridge orientation θ. Both Hradial and Hangle are defined by bandpass
filters.
Hong, Wan, and Jain [49] proposed an effective method based on Gabor filters.
Gabor filters have both frequency-selective and orientation-selective properties and
27