4.1.2 Computing Features
The following subsections describe the process used to get the
features that will be independently optimized for each
classifier. The name of the command is
listed for each step. For details on
the arguments and parameter files used, see the
Reference Manual in AppendixB. Section 4.1.3
discusses how the features are optimized for each classifier.
4.1.2.1 Make the Orientation Arrays mkoas
This command reads the fingerprint image files and extracts
the orientation array(oa). This is run on the full set of
finger prints that will be used as the S,trainingsetT^ for the
neuralnetwork
classifier.
4.1.2.2 Make the Covariance Matrix meancov
This command reads a set of oas and computes their sample mean and
sample covariance matrix. 11Is it typically run on the
full set of orientation arrays from
mkoas but could be run on just a reasonably large subset
of the training set.
4.1.2.3 Make the Eigen values and Eigen vectors eva_evt
This program reads the covariance matrix and computes the
eigen values, and the corresponding eigen vectors. The eigen
values are not needed in the training process, but may
be of theoretical
interest. The program calls a sequence of CLAPACK routines[47].
4.1.2.4 Run the Karhunen-Loe`ve Transform lintran
This command applies a specified linear transform to a
set of vectors. The transform matrix is the eigenvectors from
eva_evt. The set of vectors to which the transform
matrix is being applied is the oas files, from
mkoas, for the training fingerprints. This set of the
resulting low-dimensional Karhunen-Loe`ve(K-L)
vectors will be used as the training set for the MLP
classifier when optimizing the classifier weights.
A subset of the K-Lvectors will be used as data by
optrws (optimize regional weights command, below) to help optimize
the PNN classifier. Remember this version of
the K-L transform does not subtract the mean vector from
each feature vector. A complete version
of the K-L transform is included in the command kltran.