A comparison of the GLCM, GMRF and Gabor Energy methods over all of
the suites of texture problems is presented. These results are based
on implementations of these methods using conventional design
choices. The following implementation choices have been used.
Gabor Energy
| Program: | gaborClass |
| Wavelengths: | 2, 4 and 8 pixel |
| Angles: | 0, 45, 90, 135 degrees |
| Mask Size: | 17x17 pixels |
| Gaussian Window: | texton interpretation (sd = wavelength/2) |
| Command Line: | gaborClass -texton -lambda 2,4,8 -theta 0,45,90,135 |
GLCM
| Program: | glcmClass |
| Distances: | 1 pixel |
| Angles: | 0, 45, 90, 135 degrees |
| Re-quantization: | 32 grey levels |
| Rotation Invariance: | average features over angles |
| Command Line: | glcmClass -q 32 -af -d 1 -theta 0,45,90,135 |
GMRF
| Program: | markovClass |
| Mask: | standard 4th order symmetric |
| Command Line: | markovClass -mask std4s |
The table below gives the summary per-test-suite scores for the
implementations of GLCM, Gabor Energy and GMRF methods described
above. Two patterns emerge from this table. On microtextures, the
algorithms are ranked thus: GMRF, Gabor Energy and GLCM. On
macrotextures, the algorithms are ranked thus: Gabor Energy, GMRF and
GLCM. The "real world" textures (brodatz, grass, material, and visTex),
although containing a mix of macro- and microtextures, are
predominantly microtextures.