Using this framework, we can investigate the significance of any parameters to the overall classification results. In these experiments we have changed the number of wavelengths which contribute to the feature sets of the Gabor Energy method.
Results have been obtained for both Wide and Narrow envelopes using wavelength sets of 2,4,8 pixels and 4,8 pixels. The gaborClass program was used.
The Table below shows the per-test-suite scores for the Wide and
Narrow Gabor masks for wavelength sets of 4,8 and 2,4,8. As was shown
in a previous comparison between Wide
and Narrow envelopes, Narrow masks generally perform better on the
microtextures, with the Wide masks favoured for the macrotextures.
Furthermore, adding the extra wavelength features generally improves
the results. We would not expect much improvement for the larger
scale macrotextures with the Narrow window as this fine detail will
generally not be present.
| TEST SUITE | Wide 4,8 | Wide 2,4,8 | Narrow 4,8 | Narrow 2,4,8 |
|---|---|---|---|---|
| bomb | 0.8352 | 0.8362 | 0.8337 | 0.846 |
| bombRot | 0.8673 | 0.8697 | 0.9113 | 0.9248 |
| brodatz | 0.9102 | 0.9137 | 0.9284 | 0.9451 |
| grass | 0.864 | 0.8968 | 0.8608 | 0.89 |
| material | 0.9655 | 0.9716 | 0.9487 | 0.9678 |
| visTex | 0.8561 | 0.8859 | 0.8638 | 0.9066 |
| lattice | 0.9421 | 0.9715 | 0.8951 | 0.8919 |
| latticeRot | 1 | 1 | 0.9997 | 1 |
| mortar | 0.9454 | 0.967 | 0.8836 | 0.8758 |
| mortarRot | 1 | 1 | 0.9988 | 0.9921 |
| mortarRotS | 1 | 1 | 0.9969 | 1 |