resulting domain. We illustrate a combinatorial PCA and
neural network approach to address this problem [86].
The experimental data set consisted of PFM images of
the domains produced by an application of a number of
electrical pulses of varying length, and a total of 288 domains
were acquired for testing.
We used PCA to obtain a set of the descriptors that
characterized the individual domains. Each domain
image consisted of N × N pixels and was unfolded into
a 1D vector of N2 length. PCA eigenvectors (Figure 13b)
and corresponding weight coefficients (Figure 13c) characterized
the domain morphology. Color map of the weights
demonstrates clear differences between the domain
groups corresponding to different switching pulses
(Figure 13c). This approach illustrates use of eigenvectors
for characterization of all of the experimentally observed
features of the domain morphology, and the weights can
be used as an input parameter for the recognition by a
feed-forward neural network.
For testing of this approach, the experimental data set
was divided into training and test data sets. The PCA
over the training data set (about 15% of the domains)
was used for calculation of etalon eigenvectors, which was
used for deconvolution of the testing weight coefficients