Page 131 - Data Science Algorithms in a Week

Texture Descriptors for The Generic Pattern Classification Problem 115

As expected, the best results in Table 3 are obtained by DHoG and RFFT, i.e., by the
best descriptor and the best reshaping method.
Finally, in Table 4 the result of our best ensembles are reported and compared with
two baseline approaches: the first, named 1D, is the classification method obtained
coupling the original 1D descriptor with a SVM classifier; the second, is the best method
proposed in our previous work (Loris Nanni et al., 2012).
Included in Table 4 are results of the following “mixed reshaping” ensembles, which
are designed as follows:

MR1= 2×RCW + RRS (i.e., weighted sum rule between RCW and RRS)
MR2= 2×RCW + RRS + RDCT + RFFT
MR3= (RSHOG + RSWave) + 2 × (FFTHOG + FFTWave) (Xy means that the reshaping

method named X is coupled with the texture descriptor named Y)
MR4= MR2 + 2×1D
MR5= MR3 + 2×1D

Before fusion, the scores of each method are normalized to mean 0 and standard
deviation 1.
Table 4 includes the performance of the best ensemble proposed in our previous work
(Loris Nanni et al., 2012) that should be compared to MR2, where the fusion with 1D is
not considered.
The proposed ensembles work better than (Loris Nanni et al., 2012), except in the
two image datasets (bCI and pap). More tests will be performed to better assess the
performance when several features are available (as in bCI and pap). It may be the case
that different ensembles should be used that consider the dimensionally of the original
feature vector.
MR4 and MR5 perform similarly, with both outperforming 1D descriptors with a p-
value of 0.05 (Wilcoxon signed rank test (Demšar, 2006)). MR5 is a simpler approach,
however. This is a very interesting result since the standard method for training SVM is
to use the original feature vector. To reduce the number of parameters when MR4 or
MR5 are combined with 1D descriptors, we always use the same SVM parameters (RBF
kernel, C=1000, gamma=0.1) for MR4 and MR5 (while optimizing them for the 1D
descriptors).

CONCLUSIONS

This paper reports the results of experiments that investigate the performance
outcomes of extracting different texture descriptors from matrices that were generated by
reshaping the original feature vector. The study also reports the performance gains
offered by combining texture descriptors with vector-based descriptors.